ww/docs/build-system.md

WW package, dependency, build, and bootstrap architecture

Status: binding architecture decision

Decision date: 2026-08-10

Implementation status: specified, not yet implemented

This document selects the production architecture that replaces WW's current package driver, source-like interface protocol, work-directory reuse scheme, Make orchestration, test coordinator, and bootstrap chain. It is a greenfield decision. Migration effort is recorded only to plan implementation; it did not influence the selection.

The words MUST, MUST NOT, SHOULD, and MAY are normative.

1. Executive decision

WW will have one integrated command backed by one typed, content-addressed action graph. Language imports describe the language-package subgraph. A small, declarative ww.mod file describes only facts that source imports cannot: distribution requirements, products, generated inputs, native providers, and unusual link steps. Both descriptions lower to the same graph, scheduler, sandbox, cache, and explanation machinery. There is no general build language and no arbitrary build script. A one-directory, zero-dependency executable needs no manifest. Distributed projects use an exact lock file; fetching and updating are explicit commands, while build, run, test, documentation, and installation never access the network or rewrite project metadata. Packages produce target-specific binary export data and one object, importers consume only direct export data, and final products link a declared ordered closure. Every cached action names its compiler, target, profile, tools, sysroot, runtime, native inputs, environment, and content. The official toolchain owns these protocols and ships a pinned tool closure, but WW does not permanently own an assembler or linker.

The decisive insight is that the package graph and the build graph are not the same graph. Imports are a complete and desirable description of WW-language dependencies. They cannot honestly describe a C header tree, a host generator, an assembler, a linker script, a CRT, or a sysroot. Making imports pretend to do so hides native inputs; making every project use a programmable build framework destroys the simple ordinary path. Two small declarative front ends lowering to one action engine are simpler as a system than either lie.

The architecture is named WW Action Build in this document. That is a label, not another user-facing product: the command remains ww.

1.1 The Pike lens and attribution

Pike explicitly documented and defended these Go design choices and principles:

• language-defined imports make dependencies explicit, clear, and mechanically computable;
• unused imports and import cycles are errors; rejecting cycles improves package boundaries and independent maintenance;
• compilation speed and short edit/build cycles are primary design properties;
• a direct dependency's compiled artifact can carry the deeper public type facts needed by its clients, so an importer opens only direct dependency artifacts; and
• orthogonal, predictable concepts, fewer ways to express a construct, and a simple user experience are worth substantial implementation work.

Those points are stated in Pike's 2009 Go talk, the 2012 SPLASH article, and his 2015 simplicity talk (2009 talk, 2012 article, Simplicity is Complicated). The collective Plan 9 papers add system-wide placement of complexity, focused interfaces, and transparent text or explicitly encoded binary data. Go in Go documents one contingent case in which owning more of the toolchain simplified Go; it does not establish permanent toolchain ownership as a general Pike principle (Go in Go).

Plan 9's mk constructs the dependency graph before execution, rejects cycles and ambiguous recipes, and schedules independent work in parallel. Plan 9 used a target-specific compiler/assembler/loader family and portably encoded target object conventions (mk, mkfiles, compilers).

Modern module-path/version semantics, Minimal Version Selection, go.sum, the modern Go build cache, automatic toolchain selection, and current supply-chain policy are later Go-team designs, not principles uniquely attributable to Pike (module reference, go command, toolchain selection, toolchain rebuilding, supply-chain policy). This decision borrows some invariants from those systems but does not attribute them to Pike.

The following are this document's inferences from the documented principles:

• strict directory packages and direct binary export data are the smallest way to keep dependencies computable and compilation fast;
• a declarative native/action layer is necessary for an unmanaged language, because omitting it moves complexity into ambient shell state;
• one shared action engine is simpler than independent language and outer-build caches;
• given WW's complete-graph, frozen-build, and explainable-key hard gates, arbitrary graph-producing programs are rejected; they would require executing dependency host code before the graph is inspectable and add another permanent user programming model; and
• WW should specify a toolchain closure but should not maintain an assembler and linker forever when pinned external tools make the whole system smaller.

Modern native requirements force deliberate departures from historical Plan 9 and early Go: cryptographic source identities, lock files, explicit build/host/ target separation, sysroot and SDK identity, hostile dependency acquisition, cross-platform sandboxes, and cache-miss explanations were not their complete problem statement. WW adopts their architectural restraint, not their ambient host assumptions.

1.2 What follows from being unmanaged and native

Being unmanaged/native genuinely requires the build model to know:

• the target data layout and C ABI;
• foreign symbol spelling and visibility;
• object format, relocation model, CPU features, and assembly dialect;
• ordered objects, archives, shared libraries, linker scripts, and archive-group semantics;
• the libc, CRT, dynamic loader, SDK, runtime, and sysroot closure;
• freestanding entry and runtime policy;
• build-machine tools that generate host- or target-machine inputs; and
• ABI compatibility among compiler, runtime, native providers, and final link.

It does not follow that WW needs multiple dependency versions, semantic version ranges, a network resolver in every build, programmable build scripts, feature unification, a global namespace, or its own linker. In particular, the absence of a garbage collector says nothing about version resolution.

1.3 Binding answers to the critical questions

Question	Binding answer
Package identity	One canonical import path: the owning module identity for its root package, otherwise that identity plus / and the normalized package-relative path. The declared package name is a source qualifier, not identity.
Identity versus location/origin/version/content	All are separate. A resolver record maps identity and selected version to an origin and source-tree digest; a workspace maps identity to a local location.
Directory membership	Exactly one package per directory. Immediate selected source files belong to it. Nested directories never do.
Single-file packages	Deleted. A one-file directory package remains configuration-free.
Language dependency graph	The compiler-parsed imports alone define it. Manifest native/action edges extend the build graph, never the language graph.
Import interfaces	Direct dependencies only. Each direct .wwe contains the deep public type information needed to understand its own API.
.wwi	Deleted and replaced by deterministic, versioned binary .wwe export data. Canonical source prototypes are not an interchange format.
Package invalidation	The package action key changes when its selected own sources/generated inputs, direct export digests, compiler/toolchain, target/profile, declared environment, or protocol changes. A private transitive change does not invalidate it.
Cache key	Domain-separated SHA-256 over the canonical action record defined in section 6.6.
Cache scope	A per-user global local content store plus a project-local graph-history index. Remote import/export is explicit, never part of ordinary build.
Corruption/upgrades	Every object is rehashed on read; corrupt entries are quarantined. Tools and protocol versions are content inputs, so upgrades change keys.
Multiple versions	One selected version of a module identity. Incompatible major releases use distinct module identities ending /vN, so those identities may coexist.
Build scripts	Arbitrary scripts are forbidden. A finite declarative action may run a pinned build-machine tool in a denied-by-default sandbox.
Action authority	Exact readable inputs, writable outputs, argv, environment, execution platform, and tool closure. No network, shell, ambient PATH, clock, randomness, or undeclared filesystem access.
Acquisition	add, update, lock, fetch, toolchain fetch, cache transfer, and explicitly authorized remote observation may use the network. Artifact build/analysis and ordinary tests may not; running the finished user program is outside acquisition.
Manifest for a trivial program	No. A standalone directory containing a main package is sufficient. A manifest is required for distribution dependencies, multiple products, native providers, or generated inputs.
Configuration placement	Imports in source; identity, requirements, products, native declarations, and actions in ww.mod; exact selected closure in ww.lock; local paths in ww.work; ephemeral target/profile/output choices on the command line.
Local overrides	ww.work maps a module identity to a local source tree and records its observed digest. Imports do not change.
Target-specific files	A fixed filename-suffix selection rule; no source-level build expressions and no user-programmable selector.
OS distribution	A distributor may vendor the locked source closure or supply an exact ww.work/native-provider map. Substituted files and tools get new digests; frozen mode never silently consults the host.
Reproduction input	For supported official targets: project source, complete locked dependency-source bytes (vendor/CAS export), ww.lock, named immutable toolchain bundle, target/profile, and every declared external seed/signing input. Hashes without bytes are insufficient. Impure profiles forfeit the promise.
Assembler/linker ownership	No permanent ownership. The toolchain descriptor pins complete implementations. The current WW tools may bridge migration only.
Stage zero	One release-generated, checked-in portable C99 compiler snapshot plus a tiny declarative bootstrap plan and digest file.

1.4 Corrective protocol boundary (2026-08-11)

The first Phase 0 experiment over-scoped the protocol freeze. It turned package resolution, manifest parsing, compiler projections, action lowering, provider recursion, graph traversal, scheduling, cache policy, failure precedence, and bootstrap assertions into a declarative expression language. Its checker then implemented those operations again. That experiment is preserved as recoverable migration evidence, but it is not the production architecture.

The correction follows the separation visible in the pinned Go source. Go reads imports from source with an imports-only parse and resolves them in ordinary loader code (go/build/read.go, cmd/go/internal/load/pkg.go). The compiler writes a narrow export representation in compiler code (cmd/compile/internal/noder/writer.go), while cmd/go builds and schedules an in-memory action graph with ordinary Go functions (work/action.go, work/exec.go). Action IDs and cache storage/validation are executable hashing and storage operations, not schema programs (work.buildActionID, internal/cache). Go's cmd/dist performs concrete staged builds and checks that the final targets are not stale. Separate compiler reproducibility tests compare repeated outputs byte-for-byte, while the release process independently rebuilds and compares archives bit-for-bit (cmd/dist/build.go, reproduciblebuilds_test.go, rebuild account).

WW adopts that division, not Go's module/network/toolchain policy. Normal typed C/WW code MUST own source loading, parsing, resolution, compiler behavior, lowering, orchestration, storage, and bootstrap execution. Declarative schemas MUST describe wire representation only. Tests MUST verify executable behavior; a schema or proof-shaped record MUST NOT stand in for running it. A generator is permitted only for repetitive codec data and MUST be small, generic, deterministic, and byte-for-byte reproducible.

Phase 0 therefore freezes only WWAR framing and primitive canonical encoding; record/enum/union tags, field order, encoded defaults and record kinds; exact domain-separated digest and action-key byte formulas; compact positive and malformed-wire vectors; a small reference codec; deterministic data-only codec generation; and its repository gate. It does not freeze algorithms for deriving the represented records. Every declared record tag remains encoded, including an optional field's empty encoded_default; absence is not default insertion.

The owning implementation phases are binding:

Behavior removed from the Phase 0 experiment	Owning phase
source imports, package graph/cycles, .wwe/.wwlm, compiler export and public type/ABI projections	Phase 1
action construction, graph traversal, scheduling, CAS/cache, environment/sandbox and build failure behavior	Phase 2
manifest/lock/work/vendor text parsing, module/source resolution, fetching, source-store policy and source-tree construction	Phase 3
native/provider recursion, lowering, link-plan construction, tool adapters and platform policy	Phase 4
actual staged bootstrap, fixed-point rebuild and byte comparison	Phase 6

WW-specific guarantees remain stronger and explicit: frozen artifact builds are deterministic and offline, selections are locked, artifacts are content-addressed, cached objects are rehashed on read, and bootstrap is established by rebuilding and comparing actual bytes. At cutover there is one user-facing build path, as already required by the migration plan.

2. Normative vocabulary

Term	Exact meaning
package	The WW declarations selected from one directory, compiled together under one declared package name and one package identity.
module	A distributable, versioned source tree rooted by one ww.mod, declaring one globally stable module identity and containing zero or more packages.
project	The module or standalone package selected by the user's current command, including its declared products.
workspace	A local, non-published set of module-identity-to-directory overlays described by ww.work. It changes location, never identity.
dependency	A typed directed edge: package import, generated-input edge, tool edge, native-provider edge, runtime edge, ordered link edge, source-input edge, or bootstrap-record edge. A source-input edge is content-rooted and has no producer action; it is valid in template/final action inputs but never in GraphEdgeV1. A bootstrap-record edge selects the producer action-record or action-result record for bootstrap.compare. The edge kind is never implicit.
target/platform descriptor	A canonical architecture/platform/ABI/object/CPU/runtime description. An action labels descriptors by role: execution B, product H, and optional compiler-output T. A target triple is only a short lookup name.
artifact	An immutable byte string or canonical directory tree produced by an action and named by a content digest. Materialized files are copies or links, not the artifact's identity.
toolchain	An immutable descriptor and content closure containing the compiler, action protocol, export/ABI versions, target descriptors, resource files, runtime implementations, and pinned assembler/linker/archive tools.
sysroot	A content-identified target filesystem tree containing the exact headers, libraries, CRT objects, loader metadata, linker scripts, and SDK files exposed to target actions.
source identity	sha256 of the canonical source-tree encoding in section 4.5. It is independent of download URL and checkout path.
version	An immutable SemVer release label associated with one module identity and one source identity. It is selection metadata, not package identity.
product	A named requested result: executable, static library, shared library, object bundle, test binary, generated tree, documentation tree, or toolchain component.
action	A pure, finite build step with a typed canonical record, declared input artifacts, one execution platform, and declared output paths.
build platform (B)	The platform on which the build actions execute.
host platform (H)	The platform on which the requested product will execute.
target platform (T)	For a compiler-like product, the platform for which that product emits code. It is absent for an ordinary executable or library.

Module and package identities are slash-separated ASCII paths. They are NFC-normalized, case-sensitive, contain no empty, . or .. segment, and do not depend on filesystem case folding. A non-root package identity is written module-id/package/path; the root package identity is module-id. The selected module catalog records which module owns each package identity. If two selected modules would supply the same package identity, resolution fails rather than choosing a longer prefix.

A no-manifest invocation gives its sole root package the reserved internal identity @standalone/root; any external test package gets the reserved identity @test/<first-128-bits-of-SHA256(production-package-identity)>. These namespaces cannot be declared by a module or imported from ordinary source. The default standalone executable materializes as main, independent of directory basename. A standalone package cannot contain/import another local package or be published until ww init gives it a stable module/package identity.

3. Source and package rules

3.1 One directory, one package

A package directory contains its immediate regular source files only. The build does not follow source symlinks. Nested directories are separate packages. Every selected production source MUST begin with the same canonical package name; clause. The declared name MUST be a valid WW identifier. It need not repeat the directory leaf because identity and source qualifier are separate concepts.

The following current forms are errors after the migration:

• importing a single .ww file as if it were a package;
• placing multiple package blocks in one compilation unit;
• satisfying an unresolved import from an inline package block;
• selecting a literal source file as the build root; and
• finding packages through ordered -I search roots.

ww build ./cmd/tool selects a directory. A directory with one source file is still the smallest package and needs no extra metadata.

3.2 File membership and target selection

Production candidates are immediate regular files ending .ww, excluding *_test.ww and generated outputs. Names are byte-sorted after UTF-8 validity and normalization checks. A source symlink, duplicate normalized name, case-fold collision, or non-regular candidate is a loud error on every host.

Target variants use this only convention:

stem[+os][+arch][+environment].ww

Recognized tags come from the selected target descriptor, not from the host. Files are grouped by stem. The matching member with the greatest number of tags wins; the untagged member is the fallback. Two equally specific matches are an error. Examples are poll.ww, poll+linux.ww, and poll+linux+amd64+gnu.ww. This is replacement selection, not additive feature selection; additive code uses a distinct stem. Unknown tags are errors. There are no boolean selectors, glob expressions, or manifest-defined tag meanings.

Test variants put the same tags before the reserved suffix, for example poll+linux_test.ww; their grammar is stem[+tags]_test.ww and the identical most-specific rule applies within the test set. A production stem ending _test is reserved and rejected, preventing a tagged test from being mistaken for production source.

The selected file-name list is itself an action-key input. Therefore adding or removing a more-specific file invalidates the package even when the old files' bytes do not change.

CPU features and optimization mode always enter the compile key but do not add another WW source-selection language. WW-level specialization uses compiler intrinsics/runtime dispatch or a distinct package; CPU/float-ABI/PIC-sensitive C or assembly uses the finite native when constraints in section 8.4. This keeps ordinary source membership conventional while still making exceptional native selection exact and inspectable.

3.3 Imports, names, and resolution

The canonical forms are:

import "example.org/codec/hex";
import wire "example.org/protocol/hex";

The quoted string is the package identity. A module package may abbreviate its own module prefix with a relative import written import "./sub/path";; resolution replaces ./ with the importing module identity and normalizes the remainder without permitting ... Bare dotted imports and filesystem imports are deleted. Standalone packages cannot use relative imports.

The default source qualifier is the imported package's declared name. An alias changes only that qualifier. Two imports producing the same qualifier are an error unless one is explicitly aliased. Importing the same identity twice, resolving one identity to two sources, or resolving two selected module records to the same module identity is a loud collision error.

Resolution is exact:

1. Build the locked/workspace package catalog by joining every selected module identity with its package directories.
2. Require exactly one catalog owner for the requested package identity; zero is unresolved and two is an identity collision, even when one module prefix is longer.
3. Verify that the catalog directory exists in that module source tree and has the expected package clause.
4. Never search another root and never choose by filesystem accident.

Imports are parsed by the compiler front end, not a line scanner. Their union forms package edges, but name visibility remains file-scoped. Every imported qualifier must be used. Package cycles, including self-imports, are reported before compilation with one stable identity path through the cycle.

If an import's catalog owner is outside the importing module, that owner MUST be a direct require in the importing module's own ww.mod; availability through another dependency is not enough. This keeps distribution dependencies as explicit as package imports and prevents accidental reliance on a transitive selection. ww add creates the requirement; source is never rewritten.

The sole exception is the toolchain's intrinsic standard module, whose identity and source-tree digest are part of the selected toolchain descriptor. It is available without a manifest requirement, including to a standalone package; it is not searched from an installation directory or upgraded independently.

3.4 Visibility and internal packages

Existing exported-versus-private declaration rules survive. An internal directory segment adds one resolution rule: a package under M/P/internal/Q may be imported only by the package M/P or a package having M/P/ as a segment prefix. For M/internal/Q, the allowed root is module M and its descendants. This is checked against the catalog's owning module and package identities, not checkout paths. There are no friend lists or manifest visibility overrides.

3.5 Tests, examples, documentation, and generated WW

Only *_test.ww files are test sources. package p; tests compile with package p; package p_test; tests compile as a separate external package importing the production package normally. Test sources of dependencies are never in a consumer graph. ww test . tests one package; ww test ./... discovers package directories under the selected module, excluding hidden, underscore-prefixed, vendor, cache, and output directories. Discovery is deterministic and does not follow directory symlinks. Test compilation is cached; each selected test binary is executed on every command and independent binaries may run in parallel.

Each test invocation gets only its declared test-data mounted read-only under /data, its literal test-env, a private writable temporary directory, and the selected runner/runtime closure. Project/home/host files and network are denied. Because execution is an observation rather than an artifact action, real clock, process IDs, scheduling, and OS randomness may be exposed and are recorded as runner capabilities; their output is never cached or part of byte reproduction.

The same-package test action uses the production package identity with the non-importable action variant same-test; it compiles production and test sources together so private names remain visible. An external test uses the reserved @test/<128-bit-production-identity-digest> package identity and has a normal direct import of the production package. Neither identity can collide with or be imported by published source.

An example is an ordinary package or named product under examples/; it has no special dependency semantics. Documentation is derived from source comments and .wwe declarations, not by compiling examples during an ordinary build.

Generated WW source MUST be declared as an output of an action and as a generated input of exactly one package. It uses the suffix .wwgen, does not appear in the source tree, may not contain import or package clauses, and therefore cannot discover new graph edges after graph construction. It may refer to built-ins and declarations already in its owner package. A generator needing another package must have a checked-in owner file that imports it. This restriction keeps the complete package graph inspectable before executing generators.

A generated fragment may contain a foreign declaration only when the owning package clause already names its provider slot and the declaration explicitly names that slot. It cannot add a native requirement. After generation, the compiler verifies imports/package clauses are absent and the observed foreign slots exactly match the predeclared set; mismatch is a generator/protocol error.

4. Dependency distribution model

4.1 When metadata is required

A standalone one-directory executable with no non-toolchain dependency builds without metadata. ww init MODULE creates a module when the program needs a stable import identity, distribution dependencies, multiple packages/products, generated inputs, or native declarations.

A module tree contains exactly one ww.mod at its root; nested manifests are an error. A workspace composes separate module roots instead of nesting ownership.

ww.mod is declarative UTF-8 data. It is not WW code: it has no expressions, variables, imports, include files, macros, loops, user functions, or host conditionals. Strings use JSON escaping; lists preserve source order and record key order is non-semantic. Unknown fields are errors unless a later manifest schema is explicitly selected.

A minimal module is:

ww-manifest 1
module = "example.org/hello"
language = "1"
toolchain = { id = "ww.org/toolchain", minimum = "v1.4.0" }

require "example.org/codec" {
  minimum = "v1.2.3"
  source-index = "https://example.org/codec/.well-known/ww-source"
}

The complete set of top-level clause kinds in schema 1 is require, product, package, action, and native. Global scalar keys are only module, language, and toolchain. require has required minimum and optional credential-free source-index; the latter maps origin without changing identity. The schema tables in section 6.9 close all remaining fields. There is deliberately no general [settings] escape hatch.

The root main package is the default executable product, named after the module's last segment, with explicit normalized linkage dynamic and runtime hosted. Libraries need no product declaration to be imported. Additional or non-default outputs are explicit:

product "inspect" {
  kind = "exe"
  root = "cmd/inspect"
  linkage = "dynamic"
}

4.2 Selection rule

Versions are vMAJOR.MINOR.PATCH SemVer labels with the usual prerelease order. A requirement is one minimum version, never a range. Selection chooses the greatest minimum requested for each module identity over the complete transitive closure and repeats until stable. This is intentionally the small, monotonic part of Minimal Version Selection, not every behavior of the Go module command. The selected result is written exactly to ww.lock by add, update, or lock; build never resolves a newer version.

Exactly one version of a module identity is selected. A backward-incompatible major version N >= 2 MUST declare a module identity ending /vN, and imports name that identity. Consequently incompatible releases may coexist as distinct identities without an aliasing version resolver. Two versions of the same identity cannot coexist.

There are no feature sets, optional-dependency activation, target-dependent version constraints, upper bounds, wildcard versions, or dependency-wide configuration unification. Target variation belongs in source selection and native-provider declarations after one source closure is selected.

The selected source manifests also contribute one minimum for the same toolchain identity. The root lock chooses one exact installed/catalog version at least as high as every minimum and records its descriptor/bundle digests. A different toolchain identity or unsupported language/export/runtime protocol is an error; SemVer alone never overrides protocol compatibility. Dependency manifests do not pin the consumer to their development compiler, while an application remains exactly reproducible from its lock.

4.3 Lock file

ww.lock is generated, canonical, and committed for applications and toolchains. Published libraries SHOULD commit it for their own tests, but consumers resolve from ww.mod requirements. A lock record is:

ww-lock 1
root-manifest = "sha256:9c..."
toolchain "ww.org/toolchain" {
  version = "v1.4.0"
  descriptor = "sha256:31..."
  bundle = "sha256:80..."
  origin = "https://dist.wwlang.org/toolchain/v1.4.0/"
}
module "example.org/codec" {
  version = "v1.2.3"
  origin = "https://example.org/codec/.ww/v1.2.3.tar.zst"
  archive = "sha256:4a..."
  tree = "sha256:f7..."
  manifest = "sha256:55..."
  signature = "ed25519:key-id:base64..."
  provenance = "https://example.org/codec/.ww/v1.2.3.intoto.jsonl"
}

Records are sorted by identity. Required semantic fields are version, immutable origin, archive digest, canonical source-tree digest, and manifest digest. Signature and provenance are optional records whose verification policy is configured by the user or distributor; hashes are never optional. A signature, when present, covers schema, module identity, version, tree digest, and manifest digest. Lock files never contain local overlay paths or credentials.

4.4 Acquisition and network policy

ww add M@V discovers M by the HTTPS convention https://M/.well-known/ww-source, unless --from=URL or a user-configured longest-prefix source map supplies an index. The index returns immutable archive locations and signed digest records. Redirects and the final URL are recorded. The downloaded module manifest MUST declare exactly M; the signed index/lock record, not source text, binds V to its tree digest. Private indexes use the same protocol and obtain credentials from the fetch command's credential helper; credentials never enter build actions or lock files.

(module identity, version) is immutable: observing two signed tree digests for the same pair is an equivocation error recorded in the source store, never an automatic replacement. Yank metadata may prevent new selection but cannot alter or invalidate an already locked digest.

Only these operations may initiate network requests:

• ww add, ww update, and ww lock while selecting metadata;
• ww fetch --locked while materializing the already locked source/toolchain closure;
• ww toolchain fetch for an explicitly named toolchain; and
• explicit ww cache pull and ww cache push; and
• explicitly authorized ww observe remote execution, which is not a build, test-build, or cached artifact action.

The build/analysis phases of ww build, run, test, doc, install, graph, and explain deny network access even when an input is absent. They report the missing source digest and the exact ww fetch --locked command. They never modify ww.mod, ww.lock, or ww.work. --frozen additionally requires those files to be present, canonical, mutually consistent, and unchanged by selection. --offline is an explicit assertion of the already mandatory no-network build policy.

4.5 Source identity and storage

A module source tree contains only directories and regular files; symlinks, devices, sockets, FIFOs, absolute paths, .., duplicate normalized paths, and case-fold collisions are rejected. Its identity is:

SHA256("ww-source-tree-v1\0" ||
       for each byte-sorted relative path:
         LP(path) || type || executable-bit || LP(SHA256(file-bytes)))

LP(x) is an unsigned 64-bit big-endian byte length followed by x. Directory entries are included with type dir; regular files with type file. Ownership, timestamps, archive compression, checkout path, and non-executable permission bits are excluded. Archives are checked both against their blob digest and the unpacked tree digest. The source store is immutable and keyed by tree digest.

4.6 Workspaces, vendoring, and distributors

ww.work is local, declarative, and normally uncommitted:

ww-work 1
use "example.org/codec" {
  path = "../codec"
  expect = "sha256:f7..."
}

An overlay replaces only the location for the named module identity. The module at that path must declare the same identity. Its current canonical tree digest is an action input; expect makes accidental drift loud but may be updated by ww work sync. No import or lock identity changes.

ww vendor materializes every locked module under vendor/sha256/<tree> and writes a canonical identity-to-tree vendor/index.wwv. Frozen builds may select that source store with --vendor; the vendored bytes must match the lock. There is no flattened import tree and no rewritten import statement.

An operating-system distributor has three honest options: ship this vendor store, prefill WW's immutable source store, or provide an exact workspace/source map to distro-owned trees. Native system libraries are substituted only through the explicit provider mechanism in section 8.8. A mutable /usr lookup is an impure system profile, is local-cache-only, and is rejected by frozen builds.

4.7 Closed metadata grammar

ww.mod, ww.lock, ww.work, vendor indexes, and toolchain/native descriptors share this lexical grammar; each schema separately closes its allowed headers, clauses, fields, value types, and cardinalities:

document   = header newline { statement } EOF
header     = schema-name SP unsigned
statement  = assignment | clause
assignment = key ws "=" ws value ws newline
clause     = key ws string ws "{" newline
             { assignment } "}" ws newline
value      = string | unsigned | boolean | list | record
list       = "[" ws [ value { ws "," ws value } [ ws "," ] ] ws "]"
record     = "{" ws [ pair { ws "," ws pair } [ ws "," ] ] ws "}"
pair       = key ws "=" ws value
key        = ALPHA { ALPHA | DIGIT | "_" | "-" }
string     = JSON-string-with-valid-UTF-8
unsigned   = "0" | ("1"…"9" { DIGIT })
boolean    = "true" | "false"
ws         = { SP | TAB | newline | comment }
comment    = "#" { any-character-except-newline }

schema-name is exactly ww-manifest, ww-lock, ww-work, ww-vendor, ww-toolchain, ww-native-map, ww-native-sidecar, ww-install, or ww-bootstrap. A clause body contains assignments only, so nesting cannot grow into a language. Duplicate keys, duplicate singleton clauses, invalid UTF-8, unknown fields, integer overflow, and a comment marker inside an unclosed string are errors.

Whitespace, comments, assignment order, record-key order, and clause order where the schema declares identity keys are non-semantic. List order is semantic. Parsing produces a typed record whose canonical semantic encoding is WWAR(record), not the original text. Its semantic digest is the applicable kind/schema-bound record_id from section 6.6. ww fmt writes two-space canonical text; generated lock/vendor files MUST already equal that rendering in frozen mode.

5. Build model and graph construction

5.1 One graph, constructed before execution

For every command, ww constructs a typed graph in these deterministic phases:

1. Select the project, manifest, lock, workspace, toolchain descriptor, target, and profile. Verify their schemas, canonical identities, and content digests. Missing locked inputs are errors; this phase never fetches.
2. Enumerate selected checked-in package files by section 3.2. Ask the compiler front end to parse package clauses, imports, checked-in foreign declarations, and test metadata. Generated artifacts are known future input slots but do not yet exist. The build driver never scans source lines itself.
3. Resolve every import by identity, reject collisions/internal violations, and compute the complete acyclic package graph.
4. Add statically declared generated-input, host-tool, native-provider, toolchain, runtime, archive, ordered link, and bootstrap-record comparison edges. Match target clauses and reject zero or multiple providers. All generated output names and consumers are known here.
5. Lower nodes to the complete action-template DAG and report it. A template names every edge/input slot, tool, policy, and output, but its final key remains unresolved until every predecessor output or selected record content digest is known. Only after this point may a cache be read or a tool execute.
6. As verified cache results or completed predecessors resolve input artifacts, finalize ready action records/keys, query the cache, schedule misses, publish successful artifacts atomically, then materialize requested products.

No executed action may add a node, input, output, import, library, flag, or follow-up command. Native C/assembly declarations name complete header/source trees rather than learning dependencies from an ambient compiler depfile. This may conservatively rebuild for an unused header change, but the graph remains complete before execution and the key is correct.

ww graph --actions --format=json emits the graph after phase 5. Its canonical JSON contains node kind, execution platform, typed input/output slots, incoming edge kinds, target/profile/toolchain digests, and the exact ordered link plan. A resolved node also has its key/cache status; otherwise it has key: "pending", cache: "unknown", and a byte-sorted waiting-on list. It contains logical paths only. A collision, cycle, missing provider, undeclared target, or unresolved tool fails graph construction even if a stale cache entry might otherwise satisfy the product.

5.2 Built-in action kinds

Schema 1 has this closed set of semantic action kinds:

• ww.package: compile one package to export data, object code, and link metadata;
• ww.init: synthesize one deterministic retained package-initialization dispatcher from precomputed package link metadata;
• native.compile: compile one declared C or assembly source unit;
• archive: construct a static library from an ordered object list;
• link: construct an executable or shared library from an ordered link plan;
• generate: execute one declared build-platform tool;
• doc: render a documentation tree from sources and export data;
• bootstrap.compare: compare canonical stage outputs and manifests.

Adding an action kind changes the action schema. There is no generic rule engine, phony target, implicit suffix rule, command-string target, or shell recipe. An archive is made only for an explicit static-library product or native provider; WW packages are not automatically wrapped in one-member archives.

run and test.run are non-cacheable invocation nodes, not semantic artifact actions. test.run consumes the declared data/environment/runner and uses the test sandbox above on every request. ww run, after its network-denied build, launches the user's program with the user's runtime authority/environment unless --sandbox is explicitly requested; that execution still cannot affect a build cache entry. Both store exit status/logs only as observations. Materialization/install is a third category: a request-local side effect consuming an immutable artifact/install manifest. Neither category can satisfy or poison an artifact-action cache entry.

The semantic graph is independent of process boundaries. An implementation MAY run compiler workers in-process or in a bounded pool, but each ww.package action still has an independent canonical record and outputs. There is no required background daemon and no daemon state may affect an output.

5.3 The finite escape hatch

Unusual generation and packaging use a declarative action, not a build script:

action "protocol-bindings" {
  tool = "product:tools/schema-gen"
  platform = "build"
  inputs = {
    schema = { file = "protocol/schema.idl" }
  }
  outputs = {
    ww = { file = "generated/protocol.wwgen" }
  }
  argv = ["--input", "/in/schema", "--ww", "/out/ww"]
  env = { LANG = "C", TZ = "UTC" }
}

package "protocol" {
  generated = ["action:protocol-bindings:ww"]
}

The tool is either a named executable in the immutable toolchain or a named WW product built for H = B. /in/NAME inputs are read-only mounts and /out/NAME outputs are initially absent, exclusive writable mounts. The working directory is the empty logical /work. Output type is exactly file or tree; undeclared files fail the action. argv is passed directly, never through a shell. Input and output names are identifiers and each path appears through its fixed mount, so there is no template language.

The sandbox exposes only the declared tool closure at fixed logical /tool paths, /in, /out, the literal environment map, deterministic locale/time-zone data, and bounded CPU/memory resources. Network, process inspection, host devices, user/home directories, ambient PATH, ambient environment, wall clock, writable source, and filesystem paths outside the mounts are denied. Randomness is absent unless a declared seed artifact is mounted. A tool may spawn only executable inputs declared in its tool closure. These isolation rules apply equally to built-in compiler, C, assembler, archiver, linker, documentation, test, and bootstrap executions. Strong enforcement is required for frozen/shared-cache builds; an unsupported host must fail rather than silently weaken isolation.

This action can perform arbitrary computation over finite declared inputs, so it is sufficient for code generators, image/file-system builders, binding tools, and signing-input preparation. It cannot inspect the project and invent more work. Dependency modules may declare actions only for outputs consumed by their own packages/products; they cannot register hooks that run merely because the module is present.

Every artifact-producing action declares reproducibility = "required" in a frozen or official build. Isolation removes undeclared external inputs, but it cannot prove that arbitrary tool internals avoid PIDs, uninitialized memory, ASLR-derived values, or race-dependent output. Such variance is a tool/action defect. Official releases and first shared-cache publication of a custom generator repeat it from clean sandboxes and compare outputs. A node explicitly classified observation is never result-cached/shared and may not feed an artifact action; tests and hardware execution use that class. Impure development actions are local-only and outside the byte promise.

5.4 Build, host, and target

WW uses the conventional three-platform meaning rigorously:

B: execution platform on which build actions run
H: platform ABI of the produced artifact; executable products are intended to run here
T: output platform of a compiler-like artifact that itself is built for H

For an ordinary program T is absent. The familiar ww build --target=aarch64-unknown-linux-gnu spelling sets H; it means “build the program that runs on this target.” A compiler product may additionally set --host=H --target=T. Any generator used while producing it still executes on B; if the generator is itself WW source, its product is compiled with H = B.

Every action record carries all applicable descriptors, even when two are equal. No rule may infer H or T from the kernel running ww. Cross compilation is therefore the same graph with a different explicit host/target descriptor, not a mode that edits environment variables.

In record/JSON field names these are execution-platform, product-platform, and optional compiler-output-platform. Source suffix selection, ordinary native-provider when, sysroot, CRT, runtime, and linker selection always match the product platform H. A build-tool dependency instead has its own H equal to the parent action's B. CLI --target is only the familiar spelling for selecting the ordinary product platform; it does not rename the GNU roles internally.

5.5 Scheduling and failure

After graph construction, ready actions run in a deterministic priority order with a user-selected concurrency bound. Priority affects latency only; output bytes and link order come from records, never completion order. Independent actions may finish after another branch fails, but no dependent action starts. On the first observed failure WW stops launching work and cancels its owned in-flight actions. Which failure triggers cancellation is observational and may vary with concurrency; every concurrently observed failure is sorted by stable logical node in the report. --keep-going instead continues branches whose dependency closure remains healthy and reports all failures in that stable order. Interrupts cancel only processes owned by this invocation and leave no published partial result.

Tool stdout and stderr are captured as artifacts and streamed with node labels. Diagnostics use module-relative logical paths. --verbose may display physical mount paths separately, marked non-semantic. A successful action is published only after all declared outputs exist, have valid type/mode, are canonicalized where required, and have been hashed. A failed action is never entered in the action cache.

5.6 Atomic publication and materialization

CAS files and action-result records are written to same-filesystem unique temporary names, flushed, rehashed, then atomically renamed to their digest locations. For crash durability, WW flushes the parent directory after rename; it publishes and flushes every output before the result mapping. A concurrent publisher of the same digest verifies equality and discards its temporary file. A directory artifact is a canonical tree object whose leaves are CAS blobs. An action-result mapping is published last, so no reader can observe a result before its outputs.

Materialization is a request-local side effect outside the action-template DAG. By default requested products appear in out/<target>/<profile>/; --out and ww install --prefix change only where immutable artifacts are copied or copy-on-write reflinked. Hardlinks/symlinks are permitted only when the backing store is enforced immutable against the user and mode changes cannot affect its inode. Executable bits are set by the artifact record, never by a later ambient chmod. Replacement uses temporary siblings and atomic rename. ww clean removes materialized/project state only; ww cache gc is the separate explicit global-cache operation.

5.7 Reproducibility contract

For official supported targets, WW promises byte-identical artifacts when every artifact action satisfies reproducibility = "required" and these are identical:

• canonical project source and ww.lock;
• immutable toolchain descriptor and complete bundle;
• target descriptor and build profile; and
• all declared action inputs, including generated seeds and signing material.

The engine guarantees input isolation, logical paths, and canonical publication; repeat-build certification checks arbitrary tool determinism. The resulting promise is independent of absolute checkout, source-store, cache, output, and temporary paths; wall time, locale, process order, username, UID, and host environment are absent. Logical paths are module/package paths. Debug information uses those logical paths and fixed prefix maps. Archive metadata is canonical; timestamps and ownership are zeroed; deterministic linker build IDs derive from the link key. Official toolchains reject tools that cannot meet this contract.

A project source archive plus the complete locked dependency source bytes (a vendor/source-CAS export), its lock, the named immutable toolchain bundle, and every declared external seed/signing input is therefore a complete offline reproduction input. A lock's hashes alone cannot recreate absent bytes. Runtime behavior that depends on a shared library outside the pinned runtime/sysroot closure is not covered, and frozen official profiles prohibit such a dependency. An explicitly selected impure system profile receives no byte-identity promise, cannot publish to a shared cache, and prints every ambient input it accepted. ww verify reproducible runs isolated uncached builds under two physical roots and compares every result artifact and action manifest, not just the final executable.

6. Action records and cache protocol

6.1 Canonical encoding

Action records use WWAR 1, this deterministic byte encoding:

WWAR(record) = 0x57 0x57 0x41 0x52 | u16be(1) | value(record)
value(v)     = type:u8 | u64be(payload-length) | payload

type 0x01 bytes:  payload is the bytes
type 0x02 string: payload is valid NFC UTF-8 with no NUL
type 0x03 uint:   payload is minimal unsigned big-endian; zero is one 0x00
type 0x04 bool:   payload is exactly 0x00 or 0x01
type 0x05 list:   u32be(count) | each (u64be(value-length) | value)
type 0x06 map:    u32be(count) | each (u64be(key-length) | key-UTF-8 |
                                      u64be(value-length) | value)
type 0x07 record: u32be(count) | each (u32be(field-tag) |
                                      u64be(value-length) | value)

Record fields are strictly increasing by numeric tag. Map entries are strictly increasing by raw UTF-8 key bytes. Duplicate/out-of-order keys or tags, leading zeroes in a uint, invalid booleans/UTF-8/NFC, mismatched counts/lengths, unknown schema tags, and trailing bytes are errors. Lists preserve declared order. Schema defaults are always encoded, so no semantic field is inferred from absence. Floats, signed integers, null, and indefinite lengths do not exist.

One byte string/string is at most 2^31-1 bytes, a container has at most 2^24-1 members, and nesting depth is at most 64. A content/container-relative logical path string uses /, is relative, has no NUL/backslash, empty/./.. segment, and passes the schema's ASCII-identity or NFC-source-path rule. A field that explicitly permits . as its complete root sentinel is the sole exception. These are protocol limits, not host size_t limits. Human-readable JSON is a lossless rendering, not the hashed representation. Phase-0 golden vectors include empty/nested records, ordered lists, sorted maps, every rejection, and their complete bytes/digests. The normative empty-record vector is 57574152000107000000000000000400000000, SHA-256 138c6acb7f01e91df73cb1d9c3356d18f19d7b8eb8b0a15426bef32e515d0de0.

Each schema assigns every path field a path class. Artifact, source, generated output, install-destination, bundle-relative, sysroot-relative, and vendor-relative paths use the relative rule above. Sandbox-execution paths are path-independent absolute paths only in the closed virtual namespaces /work, /in, /out, /tool, and /data; schema-1 action working directory is exactly /work, and exact sandbox path spellings in argv are encoded. Platform-validated target-runtime paths are a separate type and may be absolute in H's namespace. Workspace locations and observation physical paths are separately typed and never enter an artifact action record or key as host locations. Absolute host paths, filesystem device/inode numbers, mtimes, cache locations, and command process IDs are invalid in action templates and final action records. A physical input enters only through a logical name, content digest, type, and semantic mode.

6.2 Required action-record fields

Every record contains, in this order:

1. WWAR schema and action kind/version;
2. language edition, compiler protocol, export protocol, object ABI, runtime ABI, manifest schema, and lock schema;
3. B, H, and optional T descriptor digests plus the expanded target fields;
4. toolchain identity, selected descriptor-slice/closure digest, compiler/backend digest, and every executable/shared/resource digest actually used by the action; the distribution bundle root/version authenticates acquisition but unused targets/tools do not invalidate this action;
5. profile fields: optimization, debug, assertions, overflow, panic, sanitizers, LTO, relocation/code model, symbol stripping, and reproducibility policy;
6. logical package/product/action identity and sandbox-virtual working directory (schema 1 exactly /work);
7. exact argument vector and a sorted literal environment map;
8. byte-sorted named inputs, each with edge kind, logical path, semantic artifact kind, semantic mode, content digest, and—where applicable—origin package identity;
9. direct export-data inputs byte-sorted by package identity for ww.package actions;
10. selected source-membership list and target-selection explanation;
11. typed native declarations: headers, objects, archives, shared libraries, sysroot, SDK, libc, CRT, dynamic loader, linker scripts and their include closures, assembler/linker/archive tools, and ABI-provider slots;
12. the exact ordered link plan, retaining archive groups, whole-archive markers, as-needed state, and repeated libraries;
13. named output paths, types, modes, and canonicalization policies; and
14. sandbox policy/version, resource bounds, and reproducibility classification.

Fields irrelevant to an action are encoded as empty values, not inferred. Native flags exist only as typed fields whose meaning is part of a tool adapter. A raw flag can be used only in a custom toolchain declaration and then its exact bytes are part of the record; the official profile has no hidden default flags.

6.3 Environment and tool discovery

No inherited environment variable is semantic. Built-in actions receive the fixed environment specified by their toolchain adapter. A declarative action receives only its env record. PATH, compiler-driver defaults, pkg-config, shell initialization, host include/library directories, and current directory are never consulted to discover an input.

User configuration may choose a cache location, concurrency, output directory, credential helper, source mirror, or display preference; these do not enter an action because they cannot alter output bytes. Choosing a toolchain, target, profile, workspace overlay, native provider, environment value, raw option, or impure system mapping can alter bytes and therefore always enters the record.

6.4 Package action inputs and invalidation

A ww.package action consumes:

• the exact selected production or test source files and generated fragments;
• their ordered membership metadata;
• only the .wwe artifacts of direct imported packages;
• the compiler/backend and toolchain resources;
• B/H/T, target descriptor, profile, language/compiler/export/object/runtime protocols, and manifest/lock schemas;
• package-specific predeclared native-provider slots, distinct from selected concrete provider declarations; and
• its literal built-in environment and sandbox policy.

It emits .wwe, one target object, and canonical link metadata. A private change in a dependency changes that dependency's object and the final link key, but not the importer's action key. A public change changes the dependency's .wwe and therefore its direct importers. If a rebuilt importer emits byte-identical .wwe, the invalidation stops there. Link-only input changes only invalidate link/archive actions; materialization merely recopies a newly selected immutable artifact when its requested result digest changes.

A package's own source bytes remain part of its action key even if a compiler could prove a change dead. A target/profile/tool/runtime ABI change creates a different key. There is no timestamp freshness shortcut and no “artifact exists” predicate.

6.5 Link action inputs

The link action consumes every reachable package/native object digest, explicit archive/shared-library digest, CRT object, runtime object, dynamic-loader choice, linker script closure, sysroot descriptor, target/profile, exact linker tool and resources, and the ordered plan. Objects are ordered by stable package identity; an explicit static product's members list controls only that archive's member order. Native archives retain declared order; repeated archives remain repeated; group and whole-archive boundaries are semantic. -L/-l token collections are not an internal representation.

The link result is cacheable. A warm identical build need not invoke the linker. Changing output/materialization path alone does not change the link key. Changing a private package implementation normally preserves importer objects but changes the final link key through that package's object digest.

6.6 Complete cache-key formula

Let R be the complete WWAR record from sections 6.2–6.5. The action key is:

K = SHA256("WW-ACTION-KEY\0" || uint64be(len(WWAR(R))) || WWAR(R))

Input entries contain the SHA-256 digest of their canonical artifact bytes/tree, not merely the producer's action key. Thus semantically identical outputs stop rebuild propagation even when their producing source/action key changed. Ordered fields remain ordered; only fields specified as maps are sorted. The domain separator and schemas prevent a digest from one protocol being reinterpreted in another.

A successful result record contains K, result-schema version, output name/type/mode/digest tuples only. Stdout, stderr, exit/diagnostic presentation, timing, worker identity, resource use, and physical paths go in a separate invocation-observation record. Thus one action key has exactly one semantic successful result even if its logs differ. Observation records may be content-addressed, but never participate in action mapping, cache hits, or reproducibility comparison.

CAS identities are type-separated. Blob bytes use their own domain; every structured object is additionally bound to its top-level record kind and schema:

blob_id = SHA256("WW-BLOB\0" || u64be(length) || bytes)

record_id(kind, schema, record) =
  SHA256("WW-RECORD\0" || u32be(kind) || u32be(schema) ||
         u64be(length(WWAR(record))) || WWAR(record))

Schema 1 reserves these top-level kind numbers: 1 tree, 2 manifest, 3 lock, 4 workspace, 5 vendor index, 6 native map, 7 platform descriptor, 8 toolchain, 9 profile, 10 action template, 11 action record, 12 action result, 13 observation, 14 graph snapshot, 15 native sidecar, 16 native ABI contract, 17 link plan, 18 export, 19 package link metadata, 20 install manifest, 21 bootstrap plan, and 22 selected-tool closure. An unknown kind is never decoded as another record. TypedDigestV1 is a record with tag 1 domain (blob or record), tag 2 algorithm (schema 1 only sha256), tag 3 record kind (0 for a blob), tag 4 record schema (0 for a blob), and tag 5 the exact 32 digest bytes. A field whose type is “typed digest” always means this record; a bare hexadecimal string is only a text rendering.

A tree record is schema 1 plus a list sorted by entry-name UTF-8 bytes. An entry is (name, kind=file|tree, executable:boolean, typed-child-digest). name is one normalized path segment. Empty directories are explicit tree children; symlinks, hardlink identity, devices, xattrs, uid/gid, mtimes, and non-executable permission bits do not exist. Duplicate normalized or case-fold-colliding names are errors. Validation recursively checks every typed child to a blob; verifying only a root digest is insufficient.

6.7 Storage, sharing, and corruption

The default cache is per-user, global across that user's checkouts, local, and private to the account:

<cache>/v1/cas/sha256/aa/bb...
<cache>/v1/actions/sha256/aa/bb...
<cache>/v1/quarantine/

The first path stores blob objects and kind/schema-bound structured records; the second maps an action key to (action-record digest, result-object digest). A small ignored project index .ww/state-v1 references the previous successful graph snapshot, whose logical nodes point to action-record/key/result digests. It roots that history until replacement so explanation can compare records and follow causes; it is disposable and never proves freshness. After explicit GC removes history, explain reports history-unavailable rather than inventing “stale.” A system-wide cache service requires authenticated isolated writers and the same signed-mapping policy as a remote cache.

On every cache read, WW verifies the requested object's digest and canonical type, decodes the mapped action record, recomputes K from it, and requires recomputed K = lookup K = ActionResultV1.tag2 plus ActionResultV1.tag3 = the mapping's typed action-record digest. It then verifies all output objects before use or materialization. A mismatch moves only that explicit entry to quarantine, removes its action mapping, reports corruption, and rebuilds. ww cache verify walks the store; ww cache gc traces retained action results and materializations. A tool upgrade changes its content/descriptor fields and cannot reuse the old key.

If two executions of one reproducibility=required action key produce different semantic result digests, WW publishes neither as an authoritative replacement, records both observations/artifact sets in quarantine, and fails with a nondeterminism diagnostic. Impure actions have no reusable action mapping.

Shared caches are opt-in explicit transports. ww cache pull imports only content-addressed objects and action mappings in an Ed25519 signed envelope over "WW-CACHE-MAP\0", cache namespace, action key, action-record digest, result digest/schema, and reproducibility/policy classification. The envelope carries a signing-key ID; configured trust policy handles rotation/revocation. Hashes prove bytes; the trusted cache signing key authorizes the asserted key-to-result mapping. All hashes are reverified. An unsigned/untrusted mapping is treated as a miss even if its referenced blobs exist. ww cache push refuses impure, non-reproducible, secret-bearing, or policy-incompatible actions. Literal secret environment values are forbidden; a required secret is a classified file input, redacted from JSON/explain, and makes the action non-shareable. Ordinary build does not contact a shared cache.

6.8 Explainability

For every node, WW retains its last local record and current record. ww explain NODE reports one of hit, not-built, missing-result, corrupt-result, policy-rejected, or key-changed. For key-changed it prints the first and, with --all, every differing typed field, for example:

codec/hex: key changed
  input direct-export example.org/base: 71… -> a4…
  caused by base: exported type Header layout changed
link hello: key changed
  package-object example.org/codec/hex: 19… -> 27…

ww explain --path NODE follows the shortest changed-input path to a source, tool, target, native provider, or policy root. --format=json exposes both WWAR renderings and field paths. Export-data differences use the normative ExportV1 semantic field/type diff; if old content was explicitly GC'd the command reports history-unavailable. Cache misses are never explained merely as “stale.”

6.9 Version-1 semantic record schemas

The following tables freeze schema-1 semantic fields and WWAR numeric tags. 1 means exactly one, 0/1 optional, * a list, and map unique string keys. Every absent optional value encodes the stated empty/default. Identity-keyed lists are byte-sorted by identity; lists marked ordered preserve source/link order. Nested records use the field tags in their named table. Enums reject unknown values rather than passing them to a tool.

Project and distribution records

Record/tag	Field	Type/cardinality	Rule/default
ManifestV1/1	schema	uint/1	1
/2	module	string/1	canonical module identity
/3	language	string/1	language edition
/4	toolchain	ToolchainRef/1	compatible ID and minimum
/5	requires	Require/*	sorted by module
/6	products	Product/*	sorted by name
/7	packages	PackageConfig/*	sorted by relative path
/8	actions	GenerateDecl/*	sorted by name
/9	natives	NativeProvider/*	sorted by name
ToolchainRef/1	id	string/1	toolchain identity
/2	minimum	string/1	minimum compatible SemVer
Require/1	module	string/1	module identity
/2	minimum	string/1	SemVer minimum
/3	source-index	string/0/1	empty means HTTPS convention
Product/1	name	string/1	unique identifier
/2	kind	enum/1	exe, static, shared, object, generated
/3	root	string/0/1	package-relative path; required except generated
/4	linkage	enum/1	dynamic, pie, static, static-pie, shared, none; kind-valid
/5	runtime	string/1	hosted default, minimal, none, or slot
/6	entry	string/0/1	empty selects typed toolchain default
/7	native	string/*	required slots, sorted
/8	linker-script	ArtifactRef/0/1	empty
/9	providers	ProviderSelection/*	sorted by slot
/10	members	string/* ordered	static/archive members; root only by default
/11	action	string/0/1	required only for generated product
/12	install-name	string/0/1	platform-validated, empty
ProviderSelection/1	slot	string/1	ABI slot
/2	use	string/1	module#native-clause
PackageConfig/1	path	string/1	normalized relative path; . root
/2	generated	string/*	sorted action:NAME:OUTPUT refs
/3	native	string/*	sorted provider slots
/4	test-data	InputDecl/*	sorted names, read-only under /data
/5	test-env	map	literal non-secret test environment

ArtifactRefV1 is permitted in declarative configuration records (including toolchain/native records) and action templates, but never in a final action record. Its tags are: 1 ArtifactNamespace (source, generated, package, toolchain, sysroot, provider-output, cas, or graph); 2 owner identity (empty only for a root source); 3 normalized logical name/path; 4 semantic artifact kind (file, tree, object, archive, shared, import-library, script, crt, loader, native-sidecar, native-abi, export, package-link, action-record, or action-result); 5 optional expected TypedDigestV1; and 6 mode (data or executable). A local source may omit tag 5 because analysis hashes it. A cas, external prebuilt, toolchain, or sysroot reference must include it. A generated/package/provider output gets its digest only from the declared predecessor output. Absolute host paths are invalid.

InputSlotRefV1, TemplateInputV1, ResolvedInputV1, built-in/template/final action-output records, and ResultOutputV1 use that same closed semantic artifact-kind enum; GenerateDecl.OutputDecl remains restricted to file or tree. In schema 1, file, object, archive, shared, import-library, script, crt, and loader require a blob digest. tree requires record kind 1, native-sidecar kind 15, native-abi kind 16, export kind 18, and package-link kind 19, each at record schema 1. Schema-1 artifact-kind values 14 action-record and 15 action-result require record kinds 11 and 12, respectively, at record schema 1. They are input-only and valid only for bootstrap.compare; they are invalid in built-in, template, or final action outputs and in ResultOutputV1. Any other digest domain, record kind, or schema is invalid kind substitution.

The graph namespace has one exact form. Its consumer is bootstrap.compare, the TemplateInputV1 edge kind is bootstrap-record, and ArtifactRefV1.tag2 is the producer logical node. Tag 3 is the Identifier selector action_record for artifact kind action-record or action_result for artifact kind action-result; tag 5 is absent. ArtifactRefV1.tag6 and TemplateInputV1.tag5 are data, TemplateInputV1.tag4 repeats the corresponding artifact kind, and TemplateInputV1.tag6 is empty. No other consumer, edge kind, selector, kind, expected digest, or mode is valid for this namespace. These inputs and edges are bijective: each graph template input has exactly one bootstrap-record GraphEdgeV1, and each such edge has exactly one graph template input. The edge's consumer node is the enclosing template node, its consumer input slot equals TemplateInputV1.tag1, its producer node equals ArtifactRefV1.tag2, and its selector equals ArtifactRefV1.tag3.

InputSlotRefV1 has tag 1 slot name and tag 2 expected semantic artifact type. A final action record contains no ArtifactRefV1: every artifact-bearing field is recursively lowered to an InputSlotRefV1. TemplateInputV1 tags are 1 unique slot name, 2 edge kind, 3 ArtifactRefV1, 4 expected semantic type, 5 semantic mode, and 6 optional origin package identity. ResolvedInputV1 tags are 1 the same slot name, 2 edge kind, 3 normalized logical name/path, 4 semantic artifact type, 5 semantic mode, 6 the resolved TypedDigestV1, and 7 optional originating package identity. It contains no producer node, producer action key, action- template digest, physical output path, or unresolved filesystem lookup. Thus all content that a native plan, link plan, source-selection record, or tool closure can read is also present exactly once in action-record tag 10 under a named slot.

GenerateDecl fields are fixed as follows: tag 1 name; 2 tool artifact/product reference; 3 platform enum (schema 1 only build); 4 TargetConstraint or empty; 5 input map of InputDecl; 6 output map of OutputDecl; 7 ordered string argv; 8 literal string environment map; 9 ResourcePolicy; 10 reproducibility enum required or impure. InputDecl is tag 1 kind (file, tree, artifact, tool), 2 logical reference, 3 optional expected typed digest, 4 semantic mode. OutputDecl is tag 1 kind (file, tree), 2 logical output path, 3 executable boolean. ResourcePolicy is tags 1 max CPU count, 2 memory bytes, 3 output bytes, 4 process count; zero selects the toolchain's recorded bound, never “unlimited.”

Record/tag	Field	Type/cardinality	Rule/default
LockV1/1	schema	uint/1	1
/2	root-manifest	typed digest/1	semantic manifest record
/3	toolchain	ToolchainLock/1	exact closure
/4	modules	ModuleLock/*	sorted identity
/5	native-map	LockedObject/0/1	empty
ToolchainLock/1…5	id, version, descriptor, bundle, origin	strings/digests	all required
ModuleLock/1	module	string/1	identity
/2	version	string/1	selected SemVer
/3	origin	string/1	exact final archive URL
/4	archive	blob digest/1	required
/5	tree	tree digest/1	required
/6	manifest	record digest/1	required
/7	signature	bytes/0/1	empty
/8	provenance	string/0/1	empty
LockedObject/1…3	origin, digest, signature	string/digest/bytes	origin+digest required

WorkV1 is tag 1 schema, tag 2 sorted Use records, tag 3 sorted local provider overrides. Use tags are module, path, expected source-tree digest. A provider override has slot, provider ID, product-platform constraint, contract digest, artifact-tree digest, and provenance in tags 1–6. VendorV1 is tag 1 schema, tag 2 lock-record digest, tag 3 sorted entries (module, version, source-tree digest, vendor-relative path) in tags 1–4. NativeMapV1 is tag 1 schema, tag 2 exact product-platform descriptor digest, tag 3 sorted provider overrides, and tag 4 signer/provenance record.

Profiles, templates, actions, results, and trees

A profile is toolchain data, not an open project map:

Tag	ProfileV1 field	Values
1	name	identity
2	optimization	0, 1, 2, 3, size
3	debug	none, line, full
4	assertions	boolean
5	overflow	trap, wrap
6	panic	abort, runtime
7	sanitizers	sorted toolchain capability IDs
8	LTO	none, thin, full
9	relocation	effective static, pic, pie
10	code-model	exact target capability ID
11	TLS default	exact target capability ID
12	strip	none, debug, all
13	reproducibility	required, impure

Tag	ActionTemplateV1 field	Rule
1	schema	1
2	kind/version	exact built-in kind or generate version
3	protocol record	language/compiler/export/object/runtime/manifest/lock; compiler protocol is distinct from compiler/backend byte identity
4	platform roles	B, H, optional T descriptor refs
5	selected tool closure	identity plus semantic closure digest/resources
6	profile	complete ProfileV1
7	logical identity/cwd	normalized identity; cwd exactly /work, path-independent
8	argv	ordered strings
9	environment	sorted literal map, no secrets
10	input slots	sorted (name, edge kind, producer/output or source ref, type, mode)
11	source selection	ordered selected membership plus reasons
12	native/link template	closed predeclared provider slots, concrete providers, and ordered link records
13	outputs	sorted name/path/type/mode/canonicalization
14	sandbox/reproducibility	exact policy/version/bounds/class

ActionRecordV1 has the same top-level tags, but tag 10 contains sorted ResolvedInputV1 records and every ArtifactRefV1 elsewhere in the template is replaced by the corresponding InputSlotRefV1. The tag-12 native/link value is therefore a resolved plan; it cannot contain an unresolved artifact reference, producer, physical path, or readable artifact locator/digest outside tag 10. ABI/layout/contract digests embedded in a referenced sidecar or contract are semantic verification values, not authority to read another object. No producer action key substitutes for a content digest.

The tag-12 native/link record separately encodes the sorted predeclared provider slots, concrete selected-provider declarations, and the optional ordered link policy or resolved plan. A ww.package action contains exactly its PackageConfigV1 native-slot list and empty concrete-provider and link-policy/ plan values; finalization copies that slot list unchanged. Concrete provider selection does not enter a package action merely because the provider satisfies one of those slots.

Finalization interns every source, direct export, package/native object, archive, shared library, header/sysroot tree, generated output, tool/resource, CRT, loader, script, and init dispatcher into exactly one named template input. For bootstrap.compare, it also interns every selected raw action record and action result as a separate named input. Finalization erases the graph namespace, producer node, and unresolved selector form. The corresponding tag-10 ResolvedInputV1 retains edge kind bootstrap-record, logical selector action_record or action_result, matching artifact kind, data mode, and the resolved typed digest; that digest is the sole authority to read the raw record. Predecessor output and selected-record digests resolve those slots lazily. The producer logical node, producer output path, producer key, and template digest are graph/provenance facts only and do not enter the consumer's ActionRecordV1 or K. Two producers that deliver the same typed bytes to the same semantic slot therefore produce the same downstream record and key.

ActionResultV1 tags are: 1 schema, 2 the 32-byte action key, 3 typed action-record digest, and 4 sorted ResultOutputV1 records. ResultOutputV1 tags are 1 unique output name, 2 semantic artifact type, 3 mode (data or executable), and 4 TypedDigestV1. ObservationV1 separately uses tags 1 schema, 2 logical invocation, 3 optional action key, 4 exit status/signal, 5 stdout blob, 6 stderr blob, 7 timing/resources, and 8 physical runner metadata; it is never an action result.

TreeV1 tags are 1 schema and 2 ordered entries. TreeEntryV1 tags are 1 name, 2 kind (file, tree), 3 executable boolean (false for tree), and 4 typed child digest. GraphSnapshotV1 tags are 1 schema, 2 logical root, 3 sorted GraphNodeV1 records, and 4 sorted GraphEdgeV1 records. GraphNodeV1 tags are 1 logical node ID, 2 typed action-template digest, 3 optional 32-byte action key, 4 optional typed action-record digest, and 5 optional typed action-result digest. GraphEdgeV1 tags are 1 consumer node ID, 2 consumer input slot, 3 producer node ID, 4 producer output name, and 5 edge kind; edges sort by that five-field tuple. Tag 4 is an ordinary producer output name except that a bootstrap-record edge uses selector action_record or action_result. That branch resolves the actual producer GraphNodeV1.tag4 or tag 5, respectively; it never selects an ActionOutputV1 or ResultOutputV1. If the selected producer tag is absent, the input remains unresolved and blocks finalization. The source-input kind is invalid in GraphEdgeV1. Non-action source inputs live only in the consumer template rather than invented graph nodes. The project index contains only its typed graph-snapshot digest.

Target, toolchain, native, interface, and handoff records

Tag	PlatformDescriptorV1 field	Rule
1	schema	1
2	name	canonical lookup name
3	arch/vendor/os/environment/object format	five exact enums
4	endian/address spaces/pointers	complete integer-width map
5	integer/data-layout	widths, alignments, aggregate algorithm
6	C ABI/data model	exact IDs and calling-convention table
7	float/variadic/name decoration	exact ABI records
8	CPU baseline/features/atomics	baseline plus required/forbidden sets
9	minimum OS/SDK	typed version record
10	TLS/unwind capabilities	sets plus defaults
11	relocation/code capabilities	supported sets plus defaults
12	executable/shared/page rules	typed object-format rules
13	hosted policy	hosted/freestanding plus capability set
14	object/runtime ABI protocols	exact IDs

TargetConstraintV1 tags 1–17 are, respectively: optional exact descriptor digest; arch; vendor; OS; environment; object format; hosted; C ABI; data model; float ABI; CPU baseline; required feature set; forbidden feature set; minimum SDK; relocation; code model; PIC requirement. Empty scalar/set means no constraint. Matching is exactly section 8.4; no expression field exists.

ToolchainV1 tags are: 1 schema; 2 ID; 3 version; 4 the same complete protocol record used by action tag 3; 5 sorted Tool records; 6 sorted platform descriptors; 7 sorted profiles; 8 sorted link policies; 9 runtime/provider records; 10 bundle tree digest/signature provenance. A Tool is (name, bundle-relative path, executable blob digest, ordered dynamic tool dependencies, resource-tree digests, adapter record) tags 1–6. A link policy is LinkPolicyV1: tag 1 product-platform descriptor; 2 product kind; 3 linkage; 4 profile constraint; 5 runtime selector (hosted, minimal, none, or an exact provider slot); 6 one ordered link-policy token template; and 7 output ABI/install policy. Policies sort by the five-field selection key and a zero/multiple match is an error. CRTs and scripts are ArtifactRefV1 tokens, not basenames. A dynamic loader is one restricted dynamic-loader token holding both its artifact and runtime interpreter path; PE/COFF platform-image policy is an ordered provider token rather than a fabricated loader artifact. SelectedToolClosureV1 deterministically projects only the relevant tools/resources/platform/profile/policy into tags 1–7; that projection—not unrelated bundle members—is action-key input.

Native artifact and ABI subrecords

Compact manifest paths are lowered to ArtifactRefV1 before WWAR encoding and then to action input-slot references before execution. The native records are:

Tag	IncludeTreeRefV1 field	Rule
1	tree	ArtifactRefV1 of kind tree
2	class	quote, user, system, or framework
3	subdirectory	normalized tree-relative path; . default

The provider's include list is ordered because header search order is semantic. The same tree may occur more than once with another class or subdirectory.

Tag	NativeSourceV1 field	Rule
1	name	unique provider-local identity
2	source	ArtifactRefV1 of kind file
3	language	exact toolchain capability ID, such as c11 or gnu-assembly
4	preprocessing	none or c-preprocessor
5	include-indices	ordered indexes into the provider include list; empty means all
6	defines	sorted literal macro map; duplicates with provider defines error

Target, profile, relocation/PIC/code/TLS policy, tool, and dialect adapter come from the enclosing native.compile record. Raw source flags do not exist.

Tag	PrebuiltObjectV1 field	Rule
1	name	unique provider-local identity
2	object	ArtifactRefV1 of kind object
3	sidecar	ArtifactRefV1 of kind native-sidecar
4	contract	ArtifactRefV1 of kind native-abi

Tag	ArchiveV1 field	Rule
1	name	unique provider-local identity
2	archive	ArtifactRefV1 of kind archive
3	sidecar	archive-level ArtifactRefV1 of kind native-sidecar
4	members	ordered ArchiveMemberV1 list in physical order
5	contract	ArtifactRefV1 of kind native-abi

ArchiveMemberV1 tags are 1 member name, 2 member blob TypedDigestV1, and 3 object-sidecar ArtifactRefV1. Duplicate names are legal only at distinct positions; member order is never sorted. The referenced archive sidecar records the ordered member-sidecar record digests as well as the physical member facts.

Tag	SharedImportLibraryV1 field	Rule
1	name	unique provider-local identity
2	kind	elf-shared, macho-dylib, or pe-import
3	link-artifact	shared object/dylib/import-library ArtifactRefV1
4	link-sidecar	ArtifactRefV1 of kind native-sidecar for tag 3
5	runtime-identity	exact SONAME, install-name, or DLL name
6	runtime-artifact	exact deployable shared object/dylib/DLL ArtifactRefV1
7	runtime-sidecar	ArtifactRefV1 of kind native-sidecar for tag 6
8	contract	ArtifactRefV1 of kind native-abi
9	runtime-requires	sorted NativeRuntimeRequirementV1 list

ELF and Mach-O tags 3 and 6 may resolve to the same bytes. For PE, tag 3 is the import library and tag 6 its matching DLL. A platform image still supplies tag 6 as a content-identified artifact within that image.

NativeProviderV1 tags are therefore: 1 name; 2 provided slot; 3 TargetConstraintV1; 4 ordered IncludeTreeRefV1; 5 ordered NativeSourceV1; 6 sorted provider define map; 7 ordered PrebuiltObjectV1; 8 ordered ArchiveV1; 9 ordered SharedImportLibraryV1; 10 sorted required slots; 11 ordered link-token templates; and 12 an ArtifactRefV1 of kind native-abi. Every ABI contract and sidecar is an independently encoded, content-addressed record. In a resolved action, those records and every artifact field above are InputSlotRefV1 values; source/action/provider output digests live only in action tag 10. A sidecar's internal artifact digest must equal the corresponding object/archive/shared input-slot digest, and its contract digest must equal the kind-16 TypedDigestV1 of the referenced native-ABI input record. Section/layout/provenance digests inside a sidecar are verification facts, not locators from which the action may read undeclared content.

Tag	NativeABIContractV1 field	Rule
1	schema	1
2	slot	exact ABI-provider slot
3	platform	NativeABIPlatformV1
4	symbols	sorted NativeSymbolContractV1 list
5	types	sorted NativeTypeContractV1 list
6	runtime-requires	sorted NativeRuntimeRequirementV1 list
7	features	NativeFeatureContractV1
8	minimum-platform	optional PlatformVersionV1
9	code	NativeCodeContractV1

Its typed record identity is record_id(16, 1, contract) as defined in section 6.6. Every schema-1 digest identifying a complete NativeABIContractV1 is the corresponding record-domain, kind-16, schema-1 TypedDigestV1. WW computes it; a supplied digest is never accepted in place of the record. Subordinate layout, calling-convention, type-contract, and header-contract digests remain their separately specified semantic values.

Tag	NativeABIPlatformV1 field	Rule
1	object-format	exact format capability ID
2	object-class	exact class/word-size ID
3	endian	little or big
4	machine-ABI	exact architecture object ABI ID
5	C-ABI	exact C ABI ID
6	data-model	exact data-model ID
7	data-layout	canonical layout digest
8	calling-conventions	canonical convention-table digest
9	float-ABI	exact ID
10	variadic-ABI	exact ID
11	symbol-ABI	exact decoration/versioning ABI ID
12	object-ABI	exact object protocol ID
13	runtime-ABI	required WW runtime ABI ID or empty

NativeSymbolContractV1 tags are: 1 exact external name; 2 exact version or empty; 3 kind (function, data, tls, ifunc); 4 role (define, require); 5 binding (strong, weak); 6 visibility (default, protected, hidden); 7 calling-convention ID or empty; 8 canonical function/object type-contract digest; and 9 optional byte size. Symbols sort by (name,version,kind,role); duplicate keys error.

NativeTypeContractV1 tags are: 1 stable binding/header-qualified identity; 2 kind (opaque, scalar, enum, struct, union, function); 3 exposure (opaque, layout); 4 canonical target-specific layout/signature digest; 5 optional size; 6 optional alignment; and 7 optional canonical header/macro contract digest. Types sort by identity and duplicates error.

NativeRuntimeRequirementV1 tags are 1 provider slot, 2 required ABI-contract digest, 3 phase (link, load, both), and 4 optional SONAME/install-name/DLL identity. They sort by (slot,phase,runtime-identity); conflicting requirements for one slot error. NativeFeatureContractV1 tags are 1 CPU baseline or empty, 2 sorted required feature IDs, 3 sorted forbidden feature IDs, and 4 sorted atomic-capability IDs; required and forbidden sets must be disjoint.

PlatformVersionV1 tags are 1 version-family ID and unsigned 2 major, 3 minor, 4 patch, 5 revision. Versions compare lexicographically over tags 2–5 only after tag 1 equality. NativeCodeContractV1 tags are 1 PIC (any, required, forbidden); 2 sorted TLS-model IDs; 3 unwind ABI ID or none; 4 sorted personality/runtime symbols; and 5 sorted required/forbidden relocation records, each record being tag 1 capability ID and tag 2 requirement (required or forbidden).

NativeSidecarV1 is evidence, not a second contract. Its tags are: 1 schema; 2 artifact TypedDigestV1; 3 evidenced NativeABIPlatformV1; 4 sorted evidenced SectionFactV1; 5 sorted evidenced SymbolFactV1; 6 sorted evidenced RelocationFactV1; 7 evidenced sorted architecture attribute/notes map; 8 evidenced NativeMachineFactsV1; 9 native ABI-contract digest; 10 sorted NativeRuntimeRequirementV1; and 11 ProvenanceV1. An evidenced value is EvidenceV1: tag 1 enum (inspected or declared) and tag 2 the value whose type is fixed by the containing field. NativeMachineFactsV1 tags are 1 NativeFeatureContractV1 and 2 NativeCodeContractV1.

SectionFactV1 tags are 1 name, 2 format type, 3 flag set, 4 size, 5 alignment, 6 optional content digest. SymbolFactV1 tags are the nine NativeSymbolContractV1 fields plus tag 10 section and tag 11 value/offset. RelocationFactV1 tags are 1 section, 2 offset, 3 exact relocation ID, 4 symbol, 5 signed addend encoded as (negative:boolean,magnitude:uint), and 6 target section. ProvenanceV1 tags are 1 producer/tool typed digest, 2 source/build record typed digest, 3 attestation bytes, and 4 signer/policy ID. A frozen opaque input requires accepted attestation for every declared value.

LinkTokenTemplateV1 and final LinkTokenV1 share tags: 1 kind; 2 artifact; 3 provider slot; 4 string value; 5 enabled boolean; 6 validated path; 7 runtime path; and 8 ordered nested tokens. Tag 2 is ArtifactRefV1 in a template and InputSlotRefV1 in a final action. Valid nonempty combinations are exactly: object/archive/shared/linker-script/version-script/export-map/def-file use tag 2; the template-only provider kind uses tag 3; group/whole use tag 8; as-needed uses tags 5 and 8; runtime-search uses tags 4 and 6; install-name, entry, and retain use tag 4; and dynamic-loader uses tags 2 and 7. Every other field encodes its empty/default value. A final LinkTokenV1 forbids provider; provider selection and its ordered fragment are recursively expanded at that exact position before the plan is final. Section 8.6 constrains valid nesting. A toolchain LinkPolicyV1 template additionally permits splice (tag 4 is exactly product-objects, init-dispatch, native-providers, or product-controls) and script-slot (tag 2 is an optional policy-default script). Those two kinds are forbidden in project/native link declarations and in a final plan.

The selected policy template is flattened in list order. Each splice kind occurs exactly once and expands to its already computed ordered product list; the single required script-slot becomes one linker-script at the same list position using the product's explicit script when present, otherwise the policy default, and disappears only when both are empty. A product script therefore replaces, never combines with, the default. Ordinary policy tokens—including CRT objects, compiler runtime, dynamic loader, and system-provider slots—stay exactly where declared. System providers are ordered template tokens, not a sorted set. Each expands to its selected provider's concrete token fragment; dependencies expand recursively, and a provider cycle is an analysis error. Every expanded artifact, sidecar, and ABI contract is a tag-10 input. A provider such as a PE platform image that adds no linker token instead contributes its resolved contract slot to the plan's non-link policy field. After expansion no provider or splice partition remains.

LinkPlanV1 tags are 1 schema; 2 the exact product-platform/kind/linkage/ profile/runtime-policy selection record; 3 the single fully expanded ordered LinkTokenV1 stream; 4 selected linker/tool/resource slot refs; 5 selected non-link platform/runtime/provider ABI-contract and sidecar slot refs; and 6 output ABI/install policy. Every file-bearing token and verification record references action tag 10. Only this resolved flattened plan enters the link action key.

ExportV1 (the WWAR body after .wwe magic) tags are: 1 schema; 2 reader capabilities; 3 language/type protocol; 4 product-platform/C/object/runtime ABI; 5 package identity/name; 6 sorted exported-surface origin/type contract table; 7 canonical type graph; 8 sorted exported declarations/constants/foreign symbols; 9 public initialization/ABI facts; 10 public-type digest; 11 public-ABI digest. Section 7.2 defines excluded non-semantic fields. PackageLinkV1 (.wwlm) tags are 1 schema, 2 package identity, 3 platform/object/runtime ABI, 4 package-object digest, 5 defined/required foreign symbols, 6 predeclared provider slots/contracts, 7 init symbol/dependency facts, 8 install/link requirements. Tag 6 contains exactly one compiler-derived ProviderContractV1 for each package action tag-12 predeclared slot, and its sorted slot projection must equal that predeclared list. It cannot add a provider slot or graph edge.

InstallManifestV1 tags are 1 schema, 2 product identity/key, 3 sorted entries (artifact digest, mode, relative destination), 4 runtime-resolution policy, 5 sorted shared/runtime closure, 6 target/toolchain/ABI provenance. Absolute prefix is deliberately absent. BootstrapPlanV1 tags are 1 schema, 2 bootstrap-host contract, 3 ordered source refs/digests, 4 portable-C compiler/output settings, 5 recorded host-C command/tool closure, 6 stage-1 outputs, 7 production toolchain closure, 8 identical logical stage-2/3/4 action roots, 9 semantic fixed-point output roles, and 10 raw-record fixed-point roles. No clause or executable step exists beyond that closed plan.

These sections fix the architectural fields, but they do not make Phase 0 an executable specification of every reference or key transformation. Phase 0 transcribes only their wire-visible record fields, tags, enum values, union discriminants, encoded defaults, field order, record kinds, wrapper framing, and digest preimage formulas into the checked-in compact schemas. Cross-field validity, construction, resolution, projection, lowering, and failure behavior belong to the executable phase that implements them. The Phase 0 generator and golden vectors determine bytes, not future build-engine semantics. No phase may silently add a wire field, renumber an assignment, change an encoded default, or alter a frozen digest formula.

7. Interface and artifact protocol

7.1 Package outputs

Every ww.package action, including a root package, emits the same three named artifacts:

export.wwe       deterministic binary export data
package.o        one target object
link.wwlm        deterministic package link metadata

These are names inside an immutable action result, not globally meaningful filenames. The cache is keyed by digests and logical package identity, so there is no __root special case and no dotted import path used as an artifact basename. A root object and dependency object obey identical protocols.

link.wwlm declares the package object's target, object ABI, defined/required foreign symbols, required native-provider slots, initialization ordering, and runtime ABI. It attests facts/edges already present in the action-template DAG; it may not introduce a provider, dependency, or action after compilation, and a mismatch is a compiler/build-protocol error. It does not contain raw linker flags. Ordinary package objects are fed directly to the product link. An archive exists only when an explicit static-library product asks the archive action to combine its ordered declared members (default: root package only). Transitive package/native dependencies remain typed link requirements in the library's install manifest and are not silently copied into multiple archives. A deliberately self-contained archive must list every member explicitly and pass duplicate-symbol/provider checks.

Language initialization never depends on linker input order. ww.init consumes all reachable .wwlm artifacts, topologically orders initialization by package imports with byte-sorted ties, diagnoses duplicate/cyclic init facts, and emits one dispatcher object. The link plan explicitly roots that dispatcher and every referenced init symbol against section garbage collection.

7.2 WW Export Data 1 (.wwe)

.wwe is a cache/build protocol, not source text and not a long-term binary distribution promise. It starts with the eight-byte magic WWEX\0\0\0\1 and a WWAR-encoded body. The magic is fixed framing that is reconstructed and verified around the body; it does not create a second blob identity. The typed .wwe identity is record_id(18, 1, ExportV1) over that body. The body contains:

1. export schema, language edition, type-system protocol, and required reader capabilities;
2. target descriptor, C ABI, object ABI, and runtime ABI digests;
3. full package identity and declared package name;
4. a sorted table of originating package/type identities and declaration-level public ABI digests actually referenced by the exported surface;
5. a canonical type graph sufficient for type checking, layout, calling convention, and code generation of every exported declaration;
6. exported constants, variables, functions, methods, types, and explicit foreign symbols; and
7. the public initialization/ABI facts needed by an importer.

Declarations are sorted by (kind, exported name, stable overload discriminator); type-graph nodes are assigned by deterministic structural traversal. Integer and floating constants use canonical target-independent bit encodings until a target conversion is part of their type. Function parameter names, source locations, comments, unused imports, private function bodies, and declaration order are not semantic and are omitted. Documentation/source mapping is a separate optional artifact and cannot invalidate an importer.

Unmanaged layout sometimes depends on facts that are private at the source level. An exported representation therefore records size, alignment, field/base offsets, calling convention, niche/tag rules, and relevant private padding or opaque-field descriptors without exposing private names. @repr("c") types also record the exact C data model and layout algorithm version. Opaque types record only the operations and layout promises permitted to clients.

7.3 Direct imports with deep public closure

An importer opens one .wwe for each direct import and no transitive interface file. If a direct dependency's API mentions a type originating in a deeper package, its .wwe embeds a canonical deep descriptor for the portion of that type required to understand and lay out the direct API. The descriptor retains the originating package/type identity and digest of that exact referenced contract—not the originating package's entire public ABI. It does not pretend the type belongs to the middle package.

This rule gives both correctness and bounded reads:

source/package imports  -> direct .wwe inputs
direct .wwe             -> complete meaning of that direct API
link graph              -> all reachable package objects

A public change in a leaf rebuilds direct reverse dependencies. Propagation continues only while each rebuilt package's .wwe bytes change. A private leaf change never enters an importer key. This replaces source-like transitive interface prepending and its quadratic composed units.

This bounds interface opens and reparsing, not necessarily total descriptor bytes: heavily re-exported type graphs can duplicate deep descriptors. Phase 1 measures total .wwe size and structural duplication on the real library graph. Only if that is material may a later export-schema revision intern immutable per-declaration descriptors; direct-import semantics do not change.

7.4 Public and ABI digests

The .wwe content digest identifies its complete target-specific bytes. It also contains two domain-separated hashes:

• public type digest over names, types, constants, visibility, and language semantics; and
• public ABI digest over target layouts, calling conventions, exported symbol contracts, runtime ABI, and representation facts.

Compile actions normally depend on the whole .wwe content digest. Tools such as documentation may depend only on the public type digest when their action kind explicitly permits it. Link compatibility checks use the ABI digest. Digest subsetting is protocol-defined; callers cannot choose arbitrary ignored fields.

7.5 Symbols and identity

Internal WW symbols are mangled from a protocol-versioned hash of the full package identity plus declaration identity, never from a leaf name or artifact filename. Resolver rules ensure only one source/version supplies that package identity. An explicitly foreign symbol is exactly the source-declared spelling and participates in duplicate-provider checks.

Package version and distribution origin are absent from mangling because they are not identity. An incompatible major version has a different /vN module identity and therefore different WW symbols. Native C symbols do not gain this protection; their provider slots and link collision rules must reject incompatible co-selection.

7.6 Compatibility and deterministic serialization

A consumer accepts only the exact export/type/object/runtime protocol combination declared compatible by its immutable toolchain descriptor. A new optional record still requires a new export schema and reader capability. Unknown records are not silently dropped. Target descriptor and ABI mismatches are errors before a compiler or linker runs.

Because .wwe is rebuilt from locked source, WW does not need an indefinitely stable compiler-internal export format. A toolchain upgrade changes action keys and may rebuild the graph. Public native-library ABI stability is a separate, explicit provider contract. Release/bootstrap byte comparisons include .wwe, objects, link metadata, action records, and executables.

8. Native integration and cross compilation

8.1 Complete target descriptors

A target triple is a user-facing alias. Before graph construction it expands to an immutable target descriptor containing at least:

• architecture, vendor, operating system, environment, and object format;
• endianness, pointer widths/address spaces, integer/long widths, alignment and aggregate-layout rules;
• C ABI/data model, calling conventions, name decoration, variadic convention, TLS ABI capabilities/default, floating ABI, and unwind model;
• baseline CPU, required/forbidden CPU features, minimum OS/SDK version, and atomic capability;
• supported/default relocation and code models, executable/shared-library rules, and page constraints;
• hosted versus freestanding policy; and
• compatible object ABI and runtime ABI protocol identifiers.

The triple x86_64-unknown-linux-gnu is insufficient by itself to identify CPU features, sysroot, glibc, loader, or linker. Those are separate descriptor/input digests. --cpu and --feature produce a new canonical expanded descriptor; the host CPU is never probed to select target features unless the user explicitly requests the impure alias native, which is rejected by frozen/shared builds.

The descriptor supplies ABI invariants plus supported/default policy values. A profile/product selects the effective relocation, code, PIC/PIE, and TLS policy from those permitted sets; the normalized effective values appear once in the action record. A conflicting or unsupported selection is rejected, never resolved by precedence between duplicate fields.

This follows the native facts exposed by LLVM data layouts and Clang's cross compilation/toolchain documentation, while making their often-driver-selected inputs explicit (LLVM data layout, Clang cross compilation, Clang toolchain).

8.2 Toolchain closure (toolchain.wwt)

An immutable toolchain bundle has a canonical toolchain.wwt descriptor with:

ww-toolchain 1
id = "ww.org/toolchain"
version = "v1.4.0"
protocols = { language = "1", compiler = "1", export = "1", object = "1",
              runtime = "1", manifest = 1, lock = 1 }

tool "wwc"     { path = "bin/wwc",     digest = "sha256:..." }
tool "cc"      { path = "bin/clang",   digest = "sha256:..." }
tool "as"      { path = "bin/llvm-mc", digest = "sha256:..." }
tool "ld"      { path = "bin/ld.lld",  digest = "sha256:..." }
tool "archive" { path = "bin/llvm-ar", digest = "sha256:..." }

target "aarch64-unknown-linux-gnu" {
  descriptor = "targets/aarch64-linux-gnu.wwt"
  sysroot = { tree = "sha256:...", path = "sysroots/aarch64-linux-gnu" }
  runtime = "runtime:aarch64-linux-gnu@1"
  libc = "c:glibc@2.39"
  link-policies = [
    { kind = "exe", linkage = "dynamic", profile = ["debug", "release"],
      runtime = "hosted",
      tokens = [
        { kind = "object", artifact = "sysroot:lib/crt1.o" },
        { kind = "object", artifact = "sysroot:lib/crti.o" },
        { kind = "splice", value = "product-objects" },
        { kind = "splice", value = "init-dispatch" },
        { kind = "splice", value = "native-providers" },
        { kind = "splice", value = "product-controls" },
        { kind = "provider", slot = "c:compiler-rt@1" },
        { kind = "provider", slot = "c:glibc@2.39" },
        { kind = "dynamic-loader",
          artifact = "sysroot:lib/ld-linux-aarch64.so.1",
          runtime-path = "/lib/ld-linux-aarch64.so.1" },
        { kind = "script-slot", artifact = "toolchain:lib/ldscripts/elf.lds" },
        { kind = "object", artifact = "sysroot:lib/crtn.o" }
      ] }
  ]
}

Every path is bundle-relative and every executable, shared tool dependency, resource directory, built-in header tree, target descriptor, runtime, sysroot, CRT, default script, and adapter is covered by the bundle's canonical tree digest. The engine invokes exact paths and passes explicit target/sysroot/resource arguments. A tool's compiled-in search outside the sandbox cannot resolve. The compact sysroot:/toolchain: references in the example normalize to full ArtifactRefV1 records containing the individual typed digest obtained from that authenticated tree; the shorthand itself never enters WWAR.

The whole bundle digest authenticates acquisition. An action key uses the transitive selected descriptor slice/tool/resource/sysroot/runtime closure only; adding an unused target or unrelated tool to a republished bundle does not cause global recompilation. Changing any selected byte/protocol still changes the key.

Official bundles may use LLVM, GNU binutils, or another implementation per target; the architecture does not expose that choice as project semantics. Tool adapters translate WW's typed record to exact argv and declare all injected inputs. A custom bundle must do the same and pass conformance/reproducibility tests before frozen mode accepts it.

WW permanently owns the language compiler, export/object/runtime ABI protocols, target descriptor schema, action engine, and official bundle definitions. It does not permanently own the assembler, linker, archiver, C compiler, or SDK. The current w6a/w6l may serve as migration inputs, then are removed once a pinned external closure passes parity. This is the smaller long-term system.

8.3 C ABI and foreign declarations

Foreign declarations are explicit source contracts:

@abi("c") @symbol("write")
@provider("c:libc")
fn c_write(fd s32, data *u8, count usize) ssize;

@repr("c")
type Header struct { tag u32; length u16; };

The compiler checks that every type has a defined representation for the selected C ABI and records symbol, calling convention, variadic status, layout, and explicitly named provider slot in .wwe/.wwlm. @provider is mandatory for each foreign declaration (a group annotation may supply it lexically), and the owning package metadata must require that slot; WW never infers it from a symbol spelling or link position. A foreign declaration with no provider, two definitions of a strong symbol, incompatible calling conventions, or mismatched layout digest is a pre-link diagnostic where possible and a mandatory link failure otherwise.

C headers are not searched or parsed implicitly. Bindings are either checked-in WW source produced by an explicit ww bindgen c command, or a declared generate action whose inputs include the exact header trees, target descriptor, preprocessor, macro map, include roots, and binding tool. The generated result records the declared whole input-tree digests, observed include trace as non-semantic audit metadata, and C ABI digest. The trace cannot add an input; a future finer-grained scan would require a new built-in action still keyed by the complete allowed include-tree digest. No build invokes ambient pkg-config; ww native snapshot-pkg-config is an explicit, impure acquisition command that converts one selected host configuration into a reviewable native provider record and content snapshot.

8.4 Native declarations

Schema 1 uses two connected declarations. A package lists the ABI slots it requires:

package "compress/zlib" {
  native = ["c:zlib@1"]
  generated = ["action:zlib-bindings:ww"]
}

A provider declares exact target artifacts and dependencies:

native "zlib-linux-aarch64" {
  provides = "c:zlib@1"
  when = { os = "linux", arch = "aarch64", environment = "gnu",
           c-abi = "aapcs64", float-abi = "hard",
           requires-features = ["neon"],
           minimum-sdk = "linux:5.10.0.0" }
  include-trees = [
    { tree = "native/zlib/include", class = "user", subdirectory = "." }
  ]
  sources = [
    { name = "adler32", source = "native/zlib/adler32.c",
      language = "c11", preprocessing = "c-preprocessor",
      include-indices = [0], defines = {} }
  ]
  defines = { ZLIB_CONST = "1" }
  objects = []
  archives = []
  shared = []
  requires = ["c:libc"]
  link = [
    { kind = "object",
      artifact = { namespace = "provider-output",
                   owner = "zlib-linux-aarch64",
                   name = "adler32", kind = "object", mode = "data" } }
  ]
  abi = { file = "native/zlib.wwabi" }
}

Allowed provider fields are exactly provides, when, include-trees, sources, defines, objects, archives, shared, requires, link, and abi. when is a finite conjunction. It may contain exact descriptor-digest, arch, vendor, os, environment, object-format, hosted, c-abi, data-model, float-abi, cpu-baseline, relocation, code-model, and pic; requires-features/forbids-features use subset/ disjoint-set matching; minimum-sdk matches only a product platform whose declared minimum is at least that value. It has no general expression. All matching candidates are retained: identical provider/artifact digests coalesce, while multiple different matches require the product's explicit provider map rather than a specificity guess.

The textual abi = { file = PATH } form parses PATH as one complete canonical NativeABIContractV1 subdocument, encodes it as top-level record kind 16, and lowers the provider field to an ArtifactRefV1 containing that typed record digest; an inline complete record is equivalent. The local source blob and parsed contract are both hashed. The file form is not a digest assertion or a build-time include, and partial contracts are invalid.

sources entries support only toolchain-declared C language editions and assembly dialects. Each source has an exact file digest and sees only the listed include trees/defines and toolchain headers. objects, archives, and shared name content-identified prebuilt artifacts plus their target/object/ABI records. A local path is content-hashed during analysis; a literal expected digest is needed only for an external/prebuilt record. A provider may mix source and prebuilt inputs, but every produced object is its own native.compile action. There is no filesystem library search.

abi is not an opaque user assertion. It is the recomputed digest of a canonical native ABI contract containing slot, product-platform ABI/data model, calling conventions, symbol names/versions/kinds, referenced C layout/header-contract digests, required runtime slots, CPU/features, minimum SDK, and PIC/TLS/unwind requirements. Foreign bindings carry the expected contract or compatible declaration-level subset. WW cross-checks the contract against compiled objects, shared/import-library sidecars, and declared providers before link.

8.5 Assembly and object files

Assembly source declares an official external-tool dialect (gnu, llvm-integrated, or another exact toolchain capability), preprocessing mode, target, and CPU feature contract. The current w6a Plan-9-style dialect is a migration input and is not accepted after cutover. The selected assembler executable and resources are action inputs. A source for one target cannot be selected for another by extension alone.

Prebuilt objects carry a sidecar native record with content digest, object format, architecture, ABI, required CPU features, defined/undefined symbols, PIC/TLS/unwind properties, and producer provenance. WW verifies the object format/class/endian/machine header, sections, symbol table, relocations, notes, and architecture attributes against every mechanically inferable sidecar fact; the same inspection recurses into archive members and shared/import libraries. Source-level C contract, libc/runtime compatibility, and provenance are not fully inferable from object bytes, so frozen opaque prebuilts additionally need a signature/attestation accepted by policy. A missing or contradictory record is an error, not permission to ask the host linker what happens.

8.6 Static/shared libraries and ordered linking

Native link is an ordered list of typed template tokens:

{ kind = "object", artifact = "object:NAME" }
{ kind = "archive", artifact = "archive:NAME" }
{ kind = "shared", artifact = "shared:NAME" }
{ kind = "provider", slot = "c:zlib@1" }
{ kind = "group", items = [...] }
{ kind = "whole", items = [...] }
{ kind = "as-needed", enabled = true, items = [...] }
{ kind = "linker-script", artifact = "file:NAME" }
{ kind = "version-script", artifact = "file:NAME" }
{ kind = "export-map", artifact = "file:NAME" }
{ kind = "def-file", artifact = "file:NAME" }
{ kind = "runtime-search", policy = "origin-relative", path = "lib" }
{ kind = "install-name", value = "@rpath/libname.so" }
{ kind = "dynamic-loader", artifact = "file:LOADER", runtime-path = "/lib/ld.so" }
{ kind = "entry", symbol = "_start" }
{ kind = "retain", symbol = "ww_init_abcd" }

Those are the complete schema-1 project/native template kinds. provider is recursively replaced at its exact position by the chosen provider fragment; the final kinds are every listed kind except provider. The toolchain-only splice and script-slot kinds also lower away as specified in section 6.9. Every file-bearing final token resolves to a declared typed artifact and digest. Nested items contain only link tokens; repetition is represented by repeating a list entry and is never deduplicated. group contains archives or provider templates that resolve only to archives, whole contains archives only, and as-needed contains shared inputs or provider templates that resolve only to shared inputs; any other expansion or nesting is invalid. Runtime-search/install-name values are validated by the selected platform adapter; frozen bundled policy permits only relocatable origin-relative paths. Entry/retain tokens become the pinned linker's typed entry/undefined-root mechanism. Raw flags exist only inside a content-identified custom toolchain adapter.

The plan is not a set and is never alphabetically reordered. The selected toolchain policy template determines global position: it can place start CRT before the product-object splice and end CRT after runtime providers, rather than relying on one universal ordering rule. Inside product-objects, WW objects use stable package-identity order. Inside native-providers, fragments use requester-before-provider topological order, which gives libA before the libB it requires. product-controls carries the declared entry/retain/install tokens; init-dispatch carries its single retained object when needed. Cyclic static archives must be represented by one explicit group; an undeclared provider cycle is an error. Repetition, whole-archive, as-needed, export maps, and symbol-version scripts remain exact records in the flattened canonical plan.

An archive action preserves declared member order and canonicalizes header timestamps, ownership, modes, and string tables. A shared-library input includes its link-time artifact, SONAME/install-name, ABI digest, transitive runtime requirements, and deployable runtime artifact digest. Merely finding the same basename in a host directory is never equivalence.

The installation manifest also fixes runtime resolution. ELF bundled policy copies the exact shared closure under a digest-namespaced relative lib/ and uses an origin-relative RUNPATH; Mach-O uses exact @rpath/install names; Windows places named DLL artifacts in the declared application directory beside their matching import libraries. A platform-image provider may instead bind an exact loader/system tree. An OS-managed mutable shared library is an impure runtime policy: link bytes can still be recorded, but WW does not promise that execution will load a particular digest.

Linker scripts are declared content inputs. The adapter resolves/audits only file-bearing directives such as INCLUDE, INPUT, GROUP, and SEARCH_DIR for the pinned linker dialect; included files and permitted sysroot trees are in the record. The pinned linker—not WW—interprets section placement, expressions, symbols, memory regions, and target semantics inside the closed sandbox. Unresolved SEARCH_DIR, absolute host paths, and implicit default scripts are errors. GNU ld documents that scripts and archive order change link semantics; WW therefore preserves rather than abstracts them away (GNU ld scripts, GNU linker).

8.7 libc, CRT, SDK, loader, and freestanding products

A hosted platform entry names one exact sysroot/SDK and the available libc, system, compiler-runtime, and WW-runtime providers; it does not name one universal CRT sequence or loader. The selected toolchain link policy for (product-platform descriptor, product kind, linkage, profile, runtime selector) supplies the exact ordered token template containing CRTs, compiler runtime, system-provider slots, script slot, and any platform-appropriate loader contract. Compiler and linker drivers run through no-defaults adapters and receive only that declared closure, so they cannot fall back to B's /usr.

A product chooses a runtime policy:

product "kernel" {
  kind = "exe"
  root = "kernel"
  linkage = "static"
  runtime = "none"
  entry = "_start"
  native = ["freestanding:boot@1"]
  linker-script = "native/kernel.ld"
}

For an executable, linkage is exactly dynamic, pie, static, or static-pie; shared-library products use shared. The toolchain maps (product kind, linkage, profile, product-platform, runtime selector) to one exact ordered policy template; selection never ignores runtime. Loader presence follows the platform's executable rules. On an ABI with an explicit program interpreter, dynamic and pie executables must name its runtime path and exact provider artifact; static and static-pie must not. A shared library has a runtime identity and dependencies but no executable program interpreter. PE/COFF-style platforms without a separate interpreter bind the exact platform-image/loader contract through system-provider policy rather than inventing a pathname. Different product policies therefore cannot accidentally share one CRT/loader sequence.

runtime is hosted, minimal, none, or a named provider. none supplies no libc, CRT, loader, or WW runtime; compiler helper routines must be supplied by a declared provider or rejected. minimal names an exact freestanding runtime. Entry symbol, memory/linker script, relocation/code model, panic/stack policy, and any boot image action are explicit. Kernel-style targets never inherit the hosted target's defaults.

Schema 1 permits none only for object/static products and static or static-pie executables. Its selected link template must contain no CRT, dynamic-loader, libc, compiler-runtime, or WW-runtime token. minimal and named providers declare their valid product/linkage set as capabilities; hosted uses the platform's hosted set. A product/linkage/runtime tuple outside that set is rejected during graph construction. The policy's single script-slot uses the product's linker-script when present, replacing the default at the same ordered position.

The target/toolchain declares the runtime capability required by every compiler-emitted helper and language operation. During graph construction a none/minimal product is rejected if selected source operations require an unavailable allocation, panic, stack, arithmetic, TLS, unwind, or other runtime capability; this is not deferred to an unexplained undefined linker symbol.

8.8 Provider conflicts and system substitution

One link namespace may select exactly one provider digest for an ABI slot such as c:zlib@1, c:libc, or runtime:ww@1. Multiple requirements for the same slot coalesce only if they resolve to the same provider and ABI digest. Different providers, ABI major slots that export colliding unversioned symbols, and two native modules claiming the same strong symbols are loud identity-collision errors. WW never chooses whichever library appears first.

When more than one contract-compatible provider matches, the root product must select one explicitly:

product "hello" {
  kind = "exe"
  root = "."
  linkage = "dynamic"
  providers = [
    { slot = "c:zlib@1", use = "example.org/zlib#zlib-linux-aarch64" }
  ]
}

The provider ID is (declaring module identity, native-clause name), rendered with # only in metadata. Selection cannot change a dependency's required ABI contract; the chosen provider must satisfy every declaration-level contract.

Two incompatible native versions can coexist only if they use distinct provider slots and their symbols/runtime names are namespaced or versioned so the link record proves no collision. Otherwise the build must adapt one behind a wrapper, use dynamic isolation, or fail. Language-level multiple-version selection cannot solve a C global-symbol collision.

A distro system provider is a normal provider record mapping exact logical artifacts to content digests and ABI metadata. If those files live under /usr, the mapping snapshots/re-hashes them before graph construction and changes the key whenever they change. It is marked impure unless the directory tree itself is immutable and content-identified. Raw -L, -l, LD_LIBRARY_PATH, compiler defaults, and build-time pkg-config are not accepted substitutes.

A portable distributor substitution uses a closed ww-native-map 1 file:

ww-native-map 1
target = "sha256:product-platform-descriptor..."
provider "c:zlib@1" {
  use = "distro.example/native#zlib"
  contract = "sha256:..."
  artifact-tree = "sha256:..."
  provenance = "https://distro.example/provenance/zlib.jsonl"
}

ww lock --native-map=FILE records the map digest/origin in ww.lock; frozen mode accepts only that exact signed/content-verified map. ww.work may contain the same provider clause for local development, but it is an impure overlay and frozen mode rejects it. This gives OS packagers an offline substitution mechanism without changing imports or silently consulting /usr.

8.9 Cross-compilation behavior

All code-generating tools execute on B. WW/C/assembly compilation for the requested ordinary product emits H objects using only H's target descriptor, sysroot, headers, runtime, and native providers. A compiler-like product built for H may later emit T code, but no T program executes during its own build. Object headers and sidecars are checked before linking, so a host object cannot silently enter a target product.

ww test --target=H always builds target test binaries. It executes them only when H = B or the toolchain declares an explicit content-identified runner (local emulator or simulator plus immutable image) as an invocation tool. Otherwise it reports “built, not run” unless --require-run was requested, in which case it fails. The runner and its platform image are action inputs; no ambient emulator is discovered.

Remote hardware/device testing is a separate explicit ww observe test --runner=NAME operation with declared endpoint/capability authority. It may use network/devices but is a non-build observation: remote state is reported, it never populates artifact/shared caches, and it is outside byte-reproducibility claims. Ordinary ww test retains the no-network policy.

This model supports new targets without running a compiler on them: an existing host toolchain adds a target descriptor, backend, object adapter, sysroot/runtime, and native providers, then builds and tests through a declared runner or hardware step. GNU's build/host/target distinction is useful vocabulary, but WW records the complete descriptors rather than only triplets (Autoconf triplets).

9. Command-line design

9.1 The ordinary path

The default workflow is deliberately short:

ww init example.org/hello
ww build
ww run -- argument
ww test .

ww init MODULE creates ww.mod, a root main.ww only when the directory is empty, and a lock selecting the currently invoked immutable toolchain. It does not add a dependency. In an existing one-directory main package, ww build works without init; the invoking toolchain and standalone source identity are shown in verbose output.

ww build [DIR|PRODUCT] builds the default root product or one named product. The default profile is the fully specified debug profile; --profile=release selects the toolchain's immutable release profile. ww run first performs that same build, then runs only a product with H = B; arguments after -- are never interpreted by the build.

There is no command that means “build and opportunistically download whatever is missing.” If a locked source or toolchain is absent, the diagnostic names its digest and asks for ww fetch --locked.

9.2 Tests, examples, docs, and installation

# Exactly one package.
ww test ./internal/codec

# Every package below the current module root.
ww test ./...

# Compile cross-target tests and require a declared runner.
ww test ./... --target=aarch64-unknown-linux-gnu --require-run

# Build an ordinary example product and render documentation.
ww build example:examples/packet
ww doc ./... --out=out/doc

# Materialize a named release product under a prefix.
ww install inspect --profile=release --prefix=/opt/ww

Test package discovery and result reporting remain deterministic under -j N. Compilation is cached, but every selected test binary runs. An installation manifest lists every copied artifact, digest, mode, runtime dependency, and relative destination and is materialized as <product>.wwinstall. Installation never discovers libraries in the prefix and never mutates the cache artifact.

9.3 Inspecting the graph and cache

ww graph --packages
ww graph --actions
ww graph --actions --format=json > graph.json
ww explain example.org/hello/internal/codec
ww explain --path --all product:hello
ww cache verify
ww verify reproducible --profile=release

graph performs analysis but no build action. A node whose producer output does not yet exist has a known action shape and incoming edges but a pending key; its final key is computed when predecessor content digests become available. explain compares available current records with the project index and likewise does not build. These commands never fetch.

9.4 Adding, fetching, updating, freezing, and vendoring

# Query the module's conventional source index, add a minimum, and lock closure.
ww add example.org/codec@v1.2.3

# The same operation with an explicit private or non-conventional origin.
ww add corp.example/codec@v1.2.3 --from=https://packages.corp/codec/

# Materialize exactly the existing lock without changing any project file.
ww fetch --locked

# Change one direct minimum and recompute the complete lock atomically.
ww update example.org/codec@v1.4.0

# Recompute after an intentional manifest edit.
ww lock

# Materialize the exact locked source closure in the project.
ww vendor --locked

# A frozen, offline build from already present locked inputs.
ww build --frozen --offline --vendor

add and update write ww.mod and ww.lock together through temporary files only after the entire selection and integrity check succeeds. lock writes only ww.lock. fetch and vendor never change selection. Without --vendor, WW uses the immutable source store; with it, every vendor digest is reverified. add --from=URL records that credential-free index as the requirement's source-index; archive redirects/final origin remain exact lock metadata.

An ordinary non-frozen build still never updates an inconsistent lock: it fails with the exact ww lock command. Frozen mode additionally rejects workspaces, impure providers, noncanonical metadata, unlocked tools, and any source digest not named by the lock. --offline is useful in automation as an assertion but does not weaken or strengthen the normal build network denial.

9.5 Cross compilation and toolchain selection

# Select a product host platform (the familiar cross-build form).
ww build --target=aarch64-unknown-linux-gnu --profile=release

# Select exact CPU semantics rather than probing the build machine.
ww build --target=x86_64-unknown-linux-gnu \
  --cpu=x86-64-v3 --feature=-avx512f

# Build a compiler that runs on H and emits code for T.
ww build compiler:wwc \
  --host=aarch64-unknown-linux-gnu \
  --target=riscv64-unknown-none-elf

# Acquire and lock an exact toolchain before building; only fetch uses network.
ww toolchain fetch ww.org/toolchain@v1.4.0
ww lock --toolchain=ww.org/toolchain@v1.4.0
ww build --toolchain=ww.org/toolchain@v1.4.0

The selected command toolchain must match the common ID, satisfy every minimum, and equal the root lock's exact descriptor. A digest selection is exact; an ID/version selection resolves only through the lock or installed signed catalog and never during build. ww toolchain list --targets shows descriptors and sysroot/runtime digests, not just triples. ww lock --toolchain changes only the lock's exact toolchain selection when the chosen version satisfies every manifest minimum; it does not rewrite those minimums.

9.6 Deliberately absent commands/options

There is no ww build --fetch, build-time package-manager hook, raw -L/-l, global import search path, arbitrary compiler/linker environment injection, manifest evaluator, or command that installs dependencies into a mutable global source namespace. Expert escape hatches are explicit toolchains, native provider records, and finite declarative actions; all remain visible in the graph/key.

10. Toolchain and bootstrap design

10.1 The durable stage-zero seed

The smallest credible recovery seed is:

bootstrap/ww0.c          generated portable C99 snapshot
bootstrap/bootstrap.plan canonical source/tool/output plan
bootstrap/SHA256SUMS     expected seed and plan digests

ww0.c is generated at release time from the same compiler sources as the production compiler, with a deliberately non-optimizing portable C backend and the minimal support routines concatenated into one translation unit. It is not a second hand-maintained compiler and is never used in ordinary production builds. The resulting ww0 retains that portable C-emission path: it contains only enough compiler, export writer, C output, and fixed-plan execution to emit stage-1 C from the enumerated source closure and have the named host C closure compile it. It has no resolver, network client, general manifest engine, cache, test coordinator, or installation framework.

The release process regenerates ww0.c and fails if its bytes differ from the checked-in snapshot. Keeping portable source rather than four opaque host binaries permits recovery on a new architecture with a C99 implementation. The cost of the simple C emitter is accepted because it bounds and exposes the seed; it is a release backend, not another production path.

The bootstrap-host contract is narrower than “any C99”: hosted C99 with CHAR_BIT == 8, exact uint8_t/uint32_t/uint64_t, two's-complement signed integers, binary file I/O, at least 32-bit address space and the published source/ object size limits. The snapshot uses no host floating-point result or undefined signed overflow. Compile-time assertions plus a tiny I/O/integer conformance probe run before compilation. Supported implementation modes and required headers are enumerated in bootstrap.plan; an implementation outside the contract is not silently called portable.

10.2 Stage transitions and fixed point

Recovery uses these exact stages:

1. A named host C implementation compiles ww0.c to ww0. Its executable, version output, command, headers, libc, assembler, and linker are recorded in bootstrap-host.wwar; they are part of the trusted base, not silently blessed.
2. ww0 bootstrap/bootstrap.plan runs on B, emits portable C for the locked production compiler/driver, and uses that same recorded host C closure to produce stage 1 with H = B. Stage 1 must run locally; distributable cross compilers are built only after the local fixed point.
3. Stage 1 uses the pinned production target/tool closure to perform a normal frozen build of the same sources for H = B, producing stage 2.
4. Stage 2 repeats the identical logical build to produce stage 3.
5. Stage 3 repeats it to produce stage 4.
6. bootstrap.compare first requires stage-2 and stage-3 semantic output sets—compiler/driver executables, export/object/runtime artifacts, and the installed semantic bundle tree—to be byte-identical. Producer action records and provenance are excluded because stage 2 was built by stage 1 while stage 3 was built by stage 2. Once those compiler bytes converge, it requires stage-3 and stage-4 semantic outputs and raw action/result records to be byte-identical; their producer compiler digests are then equal. Each raw record is a separately named action-record or action-result input in action tag 10; comparison metadata or a digest outside tag 10 is not read authority.

Stage 1 need not equal stage 2: portable-C and production backends may generate different code. Stage 2 MUST equal stage 3 semantically, and stage 3 MUST equal stage 4 completely. A canonical stage semantic manifest lists only role/output artifact digests, never its producer. A mismatch reports the first differing output or action field and is a release failure. Two absolute roots and two concurrency levels are used for the official fixed-point job.

This fixed point proves self-consistency and path/order reproducibility; it does not by itself defeat Ken Thompson's trusting-trust attack. Official releases also perform diverse seed compilation with two independently sourced host C toolchains where available, publish both bootstrap-host records, regenerate the C snapshot from the converged WW source/compiler, and compare converged semantic outputs before signing. This is additional evidence, not a claim of formal diverse-double-compilation proof.

10.3 Trusted-computing-base accounting

The bootstrap report lists, by content digest:

• ww0.c, bootstrap.plan, and their tiny C support layer;
• host C compiler/preprocessor, headers, libc, assembler, linker, loader, and OS kernel used to create/run ww0;
• bootstrap assembler/linker/archive tools and their shared/resource closure;
• source/lock/toolchain descriptor trees; and
• the SHA-256 and signature verification implementations/keys.

That is the reproducible software/input closure. The operational trusted computing base additionally includes CPU, firmware, memory/storage behavior, kernel, and execution environment; the report names the available hardware/ firmware attestations but does not pretend a source hash removes physical trust.

Nothing is called “trusted” merely because it was found on PATH. A recovery build with unpinned system tools may establish a new local fixed point but cannot claim byte identity with an official release. An official reconstruction uses a published content-identified host/bootstrap closure and compares its advertised digests.

10.4 Toolchain distribution and selection

Each release publishes source plus locked-source-closure archive, ww0.c, lock file, platform toolchain bundles, toolchain.wwt, stage-2/3/4 semantic and action/result manifests, fixed-point report, digests, signatures, and provenance. Binary bundles are conveniences verified against the descriptor, not irreplaceable seeds. A user either selects the project-locked toolchain or an exact command-line digest; WW never downloads a new compiler while building.

Runtime ABI, export protocol, object ABI, and action-schema compatibility are declared independently. A compiler refuses a runtime whose ABI it does not support. Toolchain upgrades can retain source compatibility while intentionally invalidating all relevant action keys.

10.5 New-architecture and no-compiler recovery

For a new H satisfying the bootstrap-host contract, compile ww0.c; ww0 emits the modified production compiler as portable C and the H C compiler builds the locally runnable stage 1. Add/pin H's production object/assembler/linker closure, then reach the stage-2/3/4 fixed point. Alternatively an existing B toolchain may cross-build that H compiler, but it executes only through an explicit runner. For a new T only, no target-local compiler is required: add the backend/target descriptor, object tests, sysroot or freestanding runtime, and linker provider to an existing B/H toolchain.

When no suitable WW compiler exists, the checked-in C seed is sufficient. When no C compiler exists either, a platform must provide one previously built ww0 binary plus its exact source/digest, or bootstrap a C implementation; WW does not claim a smaller physical trusted base than the machine can execute.

11. Current-system assessment

This section records the verified baseline at commit ca2cadeb4173e8190cd4c8bcc25e7da25bcdb0bc. Line references describe that commit and are implementation evidence, not requirements for the replacement.

11.1 Current invariants

• POSIX Make selects ambient CC, AR, optional ccache, and flags; builds the Cstage tools; then uses the Cstage driver to build WWstage tools into separate work directories. Make repeats manually enumerated transitive source prerequisites for each self-hosted target (Makefile, lines 7–20 and 105–279).
• ww build creates fixed-array package nodes keyed by dotted import spelling, discovers imports with a hand-written scanner separate from the compiler parser, rejects directory-package cycles, visits dependencies in DFS postorder, and invokes packages serially. The graph is capped at 256 packages (C driver, lines 333–370 and 670–700).
• An imported directory is a separately compiled package. An imported .ww file is recursively folded into the importing unit and has no node, identity, interface, object, or link artifact of its own. Inline package blocks can also satisfy otherwise missing imports (C driver, lines 396–428 and 498–613).
• A directory package selects all immediate .ww files except *_test.ww, sorts them by bytes, and has no target-specific selection rule (C driver, lines 239–301). The build path follows source symlinks while the test coordinator rejects them.
• Resolution translates dotted imports to paths and performs a global directory-before-file search over the entry root, -I roots, and inferred source library. Thus a directory in a later root beats a file in an earlier root (C driver, lines 152–224 and 966–1003).
• Every dependency emits source-like .wwi. Every importer receives the whole transitive .wwi closure, tagged with out-of-band module comments and prepended to a composed .unit.ww; the compiler reparses that unit (C driver, lines 702–809; interface writer, lines 1–24 and 505–570).
• The root compiles without compiler -I, emits no __root.wwi, bypasses the dependency export-signature path, and receives special main handling; only dependency nodes emit interfaces. ww test also injects the resolvable test package as a synthetic root edge even when source has no such import, while runtime is an implicit link edge. Imports do not fully describe even today's complete graph (C driver, lines 1066–1083 and 1131–1146).
• .wwi records exported prototypes and direct import text but no compiler, format, target, data-layout, object ABI, or runtime ABI identity. It includes non-semantic parameter names and import spelling. The compiler's -I flag both requests interface output and changes main symbol handling (compiler entry, lines 25–56 and 83–110).
• For P reachable directory packages including root, a normal driver build launches P compiler processes, P assembler processes, and one linker. It writes P-1 dependency archives itself. The link is root object, dependency archives in reverse topological order, runtime, then separately accumulated -L and -l values, losing their original interleaving (C driver, lines 1095–1249 and 1368–1460).
• -w DIR is a caller-owned mutable reuse directory, not a cache. Freshness is exact composed-unit bytes, copied driver/compiler/assembler bytes, a text mode stamp, and artifact existence/nonzero size. The graph and units are rebuilt in memory and the final executable is relinked on every invocation. Publication through .new files with the unit committed last is usefully atomic (C driver, lines 867–952, 1016–1059, and 1110–1213).
• Without -w, a build creates and deliberately retains <output-stem>.sepwork; repeating while it exists fails. With -w, the caller must create and serialize the directory. Normal driver products are always root executables linked with the runtime; there is no library-only product path, and -S merely stops after assembly.
• ww test DIR delegates to a separate coordinator. It groups same-package package p; and external package p_test; tests from *_test.ww, excludes dependency tests, composes generated source roots, and parallelizes independent test binaries with deterministic reporting (package coordinator, lines 278–417, 508–648, and 725–1057).
• Explicit ww test FILE bypasses the directory *_test.ww classifier and accepts an arbitrarily named source root. Universal directory packages delete that distinct test mode.
• Current bootstrap is mixed C/self-hosted. Make keeps the C driver fixed, uses self-hosted compiler stages to produce ww2, ww3, and ww4, and compares ww2 == ww3 and ww3 == ww4. The self-hosted driver itself is outside that fixed-point chain. The planned four stage-zero binaries are absent; the “no C compiler” route still uses host ar (Makefile, lines 825–883; bootstrap notes).

11.2 Conflated identities and accidental behavior

This table records the baseline that the implemented slices below replaced; sections 11.6–11.18 are authoritative where they conflict with it.

Concept that must be separate	Current conflation or accident
package identity	Before the local-package slices, dotted import spelling was simultaneously graph key, module/symbol prefix, artifact basename, and link identity. The implemented loader now separates source spelling, expanded canonical identity, physical directory, and storage locator.
declared name	Before section 11.18, imported directories required their declared name to equal the import-path leaf and .wwi retained only that leaf. The implemented compiler/export path now carries the declaration independently.
filesystem location	Ordered search roots silently choose/shadow a location; the same physical directory may be compiled under two import identities, while duplicate locations for one spelling produce no collision diagnostic. Paths are lexical, not content identities.
package versus file	Directories create separate-compilation nodes; files disappear into owners. The same package syntax means two compilation models.
artifact versus identity	<dotted-path>.wwi/.s/.o/.a names artifacts; root aliases to __root, which can collide with a real import.
public versus non-semantic interface	Parameter names and AST-preserved type/import spellings influence .wwi bytes and reverse rebuilds. Imports/declarations are otherwise canonically sorted; whitespace/comments and original declaration order generally do not.
compiler interface mode versus link identity	w6c -I both emits .wwi and classifies the package as a dependency for main mangling.
native dependency versus linker search	Raw -L and -l names carry no selected file, ABI, order relationship, target, or content identity.
cache location versus cache key	The explicit -w directory is both mutable namespace and freshness state; callers must serialize it.
source root versus command UX	Help historically describes . like a basename file, while implementation stats and builds it as a directory.

Other accidental constraints include fixed 256/1024-byte name/path buffers. The compiler parser silently truncates dotted full imports beyond 255 bytes, while the C driver scanner can stop advancing and hang on an import identifier at that limit; the dynamically sized WWstage scanner differs. There is no regression test for this stage divergence. Compiler, assembler, and linker launches now use structured argument vectors in both stages. Both drivers honor exact executable paths in WW_W6C, WW_W6A, and WW_W6L and otherwise select their stage-specific sibling tools. Both drivers distinguish the package-source root selected by WW_SRCLIB from the runtime-artifact root selected by WW_LIB, and apply the same empty-value and repository/install fallbacks. The two implementations remain parallel production algorithms rather than one protocol implementation.

Build and test disagree about source symlinks. External package tests are built from a generated single-file root plus -I; an external import of a multi-file production package can resolve and fold only its canonical same-named file rather than the directory package. Test work-directory names flatten / to _, so distinct lexical paths can collide. These are consequences of routing tests around, rather than through, one package model.

make install copies only ww, wwtest, and libwcc.a, while the driver needs sibling compiler/assembler/linker tools and libwwrt.a; the installed result is not a self-contained functional toolchain outside the build tree.

11.3 Scaling, invalidation, and hidden inputs

The driver performs deterministic linear action interning but now grows every package dependency vector dynamically (section 11.14). Every package compiler reads exactly one interface per direct dependency; transitive dependencies enter only the executable archive closure. Package compiler/assembler work within one driver remains serial; Make gains parallelism only by launching independent top-level driver builds.

The observed invalidation rules are:

• a private change in a directory dependency rebuilds that package and the unconditional final link, but not importers;
• an exported change rebuilds direct importers and continues through an ancestor only while the regenerated direct-dependency export bytes change, stopping at the first byte-identical regenerated interface;
• a private change in a folded file import rebuilds its entire owner;
• a link-only option reruns the always-executed link but not package compiles;
• changing copied compiler or assembler bytes rebuilds every package; and
• nonzero corruption of .s, .o, .a, or .wwi may be accepted because content is not rehashed.

Hidden or incompletely modeled inputs include CC, AR, PATH, ccache, Make flags, compiler built-ins, assembler/linker defaults, inferred argv[0] library locations, current working directory, file mode, runtime archive, linker binary, native-library resolution, host libc/CRT/loader, SDK, CPU, target, and environment. Make does not invalidate existing C objects when the host compiler or C flags change. Persistent package workdirs now bind the exact driver executable, conservatively covering its graph, unit, archive, and commit semantics; the remaining ambient inputs are not thereby promoted into a cache protocol. WW has no target triple, sysroot, conditional source, generator, manifest, lock, source digest, or frozen/offline concept.

11.4 Measurements and disposable experiments

Measurements ran on Linux 6.12.76_1 x86-64 with eight logical CPUs, GCC 14.2.1, binutils 2.44, and no ccache. They used successful clean builds and isolated temporary source fixtures; no production migration was begun. Clean/full means use two timed samples and representative cold means three, without CPU isolation or OS-cache flushing, so they are observed baselines rather than universal performance claims.

Measurement	Result
clean make all, -j1	24.847 s mean
clean make all, -j8	15.515 s mean; only 1.60× speedup
warm make all -j8	0.02365 s mean at Make level
clean tool trace	69 package compiles/assembles in six driver workdirs; 71 w6c, 76 w6a, six w6l, 63 in-driver archives; package work inside each driver stayed serial
small / 8-package / 15-package cold driver build	0.0103 / 0.1326 / 9.6706 s
same warm driver builds	0.00814 / 0.01347 / 0.05302 s; each still linked
out/	34,464,015 bytes, 398 files
six WW build workdirs	28,609,092 bytes, 351 files; 9,954,428 bytes (34.8%) duplicate beyond the first content copy
15-package root composed unit	14 .wwi sections, 1,464,581 bytes; 62 interface insertions across graph, 1,928,898 composed-unit bytes versus 23,664 distinct interface bytes

A three-package root -> mid -> leaf fixture confirmed that root consumes both direct and transitive interfaces. A private leaf implementation edit rebuilt only leaf plus link. Adding a compatible public leaf API rebuilt all three even though mid emitted identical .wwi. A link-only -L change ran only the link.

Switching the compiler executable by byte content invalidated all packages. Appending data to a nonzero cached mid.s was not detected; the old object was reused and the build succeeded. This directly rejects artifact-presence caching.

Two builds from different absolute source roots produced the same 398 relative paths and all sampled WW-generated executables/package artifacts were byte identical. The full trees were not: 45 Cstage host objects/copied tools differed, including GCC DW_AT_comp_dir. Thus current checkout-independent byte identity holds for measured WW artifacts, not for the complete build.

Cstage and WWstage drivers building the same eight-package graph produced a byte-identical final executable and all 38 non-tool artifacts, but took 0.138 s and 0.356 s respectively. The audit also found that Cstage formerly returned success after an ambient chmod lookup failed and left mode 0644, while WWstage created mode 0755 directly. Cstage now calls chmod(2) on the exact output path and reports failure, removing that host-tool and path-splitting asymmetry.

11.5 What survives and what is deleted

The replacement retains these sound concepts:

• explicit source imports, package clauses, exported/private declarations, and loud directory-package-cycle errors, strengthened to every package after file folding (whose cycles were merely visit-deduplicated) is deleted;
• directory package boundaries, made universal rather than optional;
• same-package and external *_test.ww semantics, dependency-test exclusion, deterministic discovery/reporting, and always executing selected tests;
• deterministic sorting/serialization, atomic artifact publication, fixed-point bootstrap checks, and byte-identity tests; and
• the ordinary ww build, run, and test user experience.

The replacement deletes these concepts rather than emulating them indefinitely:

• file imports, inline multi-package units, bare dotted imports, -I search roots, directory-before-file precedence, and the __root artifact alias;
• .wwi, //ww:module wrappers, .unit.ww, transitive interface prepending, source-prototype interchange, automatic per-dependency archives, and -w;
• raw ambient -L/-l, compiler/linker/sysroot defaults, and build-time pkg-config discovery;
• separate C and self-hosted production drivers, the production Make graph, and the separate source-composing test coordinator; and
• the current Cstage bootstrap path and permanently owned assembler/linker after the generated C seed and pinned external tool closure replace them.

There will be no compatibility alias that silently translates an old import, interface, workdir, or link search into the new model.

11.6 Implemented local package slice

The first executable package slice is intentionally smaller than the final module design above. It is local, offline, and manifest-free. The supported form is:

out/bin/ww build -I /work/acme -o app /work/acme/cmd/app

Every selected source uses the existing syntax:

package main;
import lib.math;

A source import is translated from dots to path separators, expanded through the nearest eligible local vendor directory described in section 11.16, then falls back to directory lookup through the entry package's directory, explicit -I roots in command order, and the toolchain source-library root. A same-named .ww file is neither a match nor a shadow for an import, so a later root containing the directory wins over an earlier file decoy. There is no network, manifest, or imported-file fallback. Explicit single-file CLI roots retain their raw-unit compatibility path. The loader uses the compiler frontend's imports-only parser, retains every real import occurrence with its owning source file and position, byte-sorts and deduplicates the resulting canonical direct edges, interns canonical directory actions, and reports self-imports and stable cycle chains before compilation. Occurrence retention makes contextual internal and vendor checks run at every import site; it does not duplicate package actions or compiler inputs.

A directory package consists of its immediate .ww entries whose basenames do not begin . or _: regular files and symlinks targeting regular files are included under the entry name, while symlinks targeting directories are ignored. The production variant excludes *_test.ww; each variant retains byte-sorted filename order. Every selected production file must declare the same package name, but that declaration is independent of the directory name and every component of the canonical import path. A selected command directory declares package main while retaining its complete canonical import identity. A source import of a command package from a different directory is rejected; the one same-directory exception is an external main_test variant's canonical import of its production action. Two ordinary logical identities for one physical directory are rejected rather than compiled twice; section 11.16 records the deliberate exception for distinct expanded vendor routes that converge through symlinks.

Packages compile serially in dependency-first postorder. The compiler emits the existing deterministic .wwi interface for every directory-package action, including an executable root. Its primary section contains that package's byte-sorted direct imports and exported declarations. The compiler then appends byte-sorted, origin-tagged sections for only the foreign type and constant facts recursively reachable from the primary public signatures. Reachable owner-local private nominal types are carried without export: they make the export self-contained for type checking, but qualified source lookup still rejects their names. Checked fixed array dimensions are emitted as numeric type facts, so a public layout never requires exposing the private constant spelling that produced its length.

A package compilation unit contains only that package's own byte-sorted sources, deterministic //ww:module-reset separators, and sorted driver-private resolution metadata; it never contains a dependency source body. Each direct import is a separate --import <canonical-path> <dependency.wwi> compiler input, sorted by canonical path and deduplicated by the loader; a source spelling expanded through vendor additionally receives the non-dependency --import-map described in section 11.16. No transitive .wwi is passed. Origin-tagged facts inside those direct artifacts are compiler data, not source imports: a source qualifier is visible only in the source file that directly imports it. The qualifier is the imported export's declared package name, not the source spelling or import-path leaf. Private members, transitive-only qualifiers, bare values, and bare types remain compiler errors. In -c package mode the compiler parses each export independently, then coalesces repeated exported type/constant facts with the same origin, kind, and name, preserving one nominal type identity across diamonds; raw non-package w6c retains its existing one-source behavior. The source-like .wwi syntax remains a transitional export encoding pending the binary .wwe format described above, but the separate direct-input ownership boundary is live in production Cstage and WWstage compilers and drivers.

A selected production root is one normal package action. Its finalized canonical import identity tags its owner-only unit; it receives only direct exports, emits .wwi, .o, and a deterministic .a, and is compiled exactly once. The declared package name is semantic package content but never validates, shortens, aliases, or replaces that identity. After loading and identity finalization, the declaration also selects the terminal build action: package main is a command and every other valid declaration is a compile-only library. A command root receives the narrow compiler --entry flag, which controls bare main codegen, then the linker receives that root archive first, the complete reachable package-archive closure, and the runtime archive; it never receives .wwi. A non-main root receives no --entry and never enters the linker. main validates command kind but never replaces or truncates an identity such as cmd.tool. The explicit package-less single-file compatibility path has no directory package declaration to classify and remains a raw command unit; it does not participate in canonical directory-package interning. The package driver still uses the parser-only --command-package marker for command test variants that must remain ordinary archive code. Neither marker changes export identity, and imported interfaces retain independent canonical owner and declared-name records. The linkers seed main before archive selection, so the existing WWAR member protocol needs no special root object or format change.

Publication is separate from that semantic action choice. ww build -o lib.a DIR or ww build -I ROOT -o bar.a foo.bar automatically publishes a non-main root's deterministic archive and self-contained compiler export at FILE.wwi, without invoking the linker; the latter retains foo.bar.* symbols. Without -o, a non-main root and its dependencies are compiled in the selected scratch or persistent work directory and no cwd product is invented. For a command, -o continues to name the executable publication path. Output names, request order, and whether publication was requested never enter package or action identity. The historical action-selecting -p exception is removed and rejected as an unknown build flag. Assembly-only -S still stops before object, archive, publication, or link production. A literal directory is reverse-resolved through the active source roots or receives the deterministic local identity described below; its declaration can never invent or truncate that identity. Two cold builds with identical inputs are required to produce byte-identical requested products. Compiler intrinsics keep their package-mode runtime ABI independent of transitive source interfaces (for example, alloc lowers to the runtime allocator without requiring an rt.wwi compiler input).

This rule follows the pinned official Go 1.26.5 source at commit c19862e5f8415b4f24b189d065ed739517c548ba. go/build.Package stores source directory, declared name, and import path independently, and defines a command solely as a package named main (go/build/build.go, lines 436–449, go/build/build.go, lines 514–519). The Go builder's AutoAction links only main and returns the archive compile action for every other package (cmd/go/internal/work/action.go, lines 450–456). The build command invents a default output only for one main, applies an explicit -o to either AutoAction result, and otherwise builds each requested package without conflating publication and semantic kind (cmd/go/internal/work/build.go, lines 473–478, cmd/go/internal/work/build.go, lines 508–548, cmd/go/internal/work/build.go, lines 551–558).

Both WW stages represent the rule with the existing loaded declaration and stable root action index; no package, dependency, ownership, locator, traversal, or closure allocation is added. ww run adds only a command-kind requirement: a successfully loaded, cycle-free non-main root produces the deterministic package PATH is not a main package diagnostic before any compiler, assembler, or linker invocation; package and graph failures retain precedence. Before a non-main -o build invokes a producer, both stages validate the longest atomic publication spelling, FILE.wwi.new, so an incomplete archive/export pair is never caused by a late path-overflow failure. Package loading, cycle detection, and closure validation retain diagnostic precedence over this publication-only check, and -S does not validate a publication path it never consumes. Build workdir format 14 and test workdir format 13 invalidate older unit vouchers before reuse because source binding and vendor-directory identity now participate in compiler argv and persistent unit semantics. Thereafter an equivalent warm library build invokes no tools, a private dependency change stops at its unchanged export, and an export change recompiles its direct importer under the existing propagation rule.

11.7 Implemented directory package-test slice

Directory tests now enter that same local package loader and build path. The supported manifest-free forms include ww build DIR..., ww build DIR/..., ww test DIR..., and ww test DIR/..., with overlapping direct and recursive roots. Test retains its existing -run, -filter, -list, -timeout-ms, -j, -c, and -w forms. ww test -c -o test.bin DIR names the result when the selected directory has one test variant; the coordinator rejects one output name for a multi-variant or recursive test request. Explicit ww test FILE retains its compatibility path.

The test coordinator still discovers requested directories, enumerates the test package names, selects variants, executes independent binaries, and emits captured results in byte-sorted package order. It no longer concatenates a generated production/test root, resolves imports, or starts one package graph per binary or directory. Instead it sends one ordered build request containing every selected directory/variant root to the Cstage or WWstage command. Its semantic selections are only the directory and selected variant/package identities; it also carries output destinations, coordinator-private completion paths, import search roots, and the optional command-scoped work-directory policy. The command owns source selection, package loading, compiler inputs, archive construction, generated-main construction, and linking for every package variant:

• The ordinary production action selects the directory's byte-sorted non-test files. A no-test requested product uses this action directly, and every ordinary, internal-test, or external-test import of that package reuses it.
• The internal production-plus-test variant selects the byte-sorted production files followed by byte-sorted matching package p test files. It is distinct from production so its private production declarations and test declarations remain isolated to that test product.
• When production sources establish p, the external variant selects only matching package p_test files, where p is the production declaration rather than an import-path leaf. A test-only directory may establish its own one consistent test package name, matching pinned go/build classification. A real source import of the production package uses its canonical import path and is a direct edge to the same ordinary production action used by ordinary products and transitive imports; there is no external-production role or artifact.
• Generated main is a separate package action whose owner-only generated unit declares package main and imports exactly the selected variant and test support. It consumes those direct .wwi files, emits its own .wwi/.o/.a, and alone receives compiler -T --entry. Its target edge also receives the compiler-private --test-target-package <canonical-path> binding. That private canonical-path qualifier prevents a tested command declared main from colliding with generated main; it is not user alias syntax.

The authoritative directory-action identity is the triple (canonical filesystem directory, canonical ordinary import path, semantic variant). The semantic variants are production, internal production-plus-test, and external _test; generated main and the reserved test-support alias are explicit non-directory action classes. The declared package name is stored separately and is not part of canonical directory/path interning. Requested-root state, discovery role, product ordinal, output path, persistent artifact key, and discovery order never enter the triple.

A literal directory root may enter the interner before its full import spelling is known. It is provisionally interned by canonical directory and variant, and a later source import of that directory binds and reuses the provisional action. After all source discovery, but before generated-main construction or any tool invocation, each still-unbound directory is finalized by this exact algorithm:

1. For every import-resolution context that reached the directory, walk that context's roots in its normal forward precedence: the selected package's directory, explicit -I roots in command order, then WW_SRCLIB or the selected toolchain source root. A request pattern never becomes an import root and therefore cannot shorten, replace, or donate package identity.
2. Canonicalize each candidate root and require the package directory to be a strict descendant. Every relative path component must be a non-keyword WW identifier. Convert separators to dots, then resolve that relative spelling again through the complete ordered context. Accept it only if ordinary forward lookup selects the same canonical directory. Thus an earlier shadow invalidates a name inferred from a later or nested root.
3. Bind the first precedence-valid candidate from each reaching context through the command-global bidirectional interner. An identity supplied by successful logical package lookup, such as encoding.utf8, is already bound and is preserved exactly. That forward-selected identity is authoritative: reverse derivation applies only to still-unbound literal roots, so a nested active root cannot rename an explicitly resolved package.
4. If no active root can represent the directory, bind the reserved, non-source-importable identity __wwlocal.p<escaped-canonical-absolute-directory>. The escape is injective and reversible over path bytes: ASCII letters and digits are copied, _ becomes _u, / becomes _s, and every other byte becomes _xHH with lowercase hexadecimal. Source imports of __wwlocal or any of its children are rejected, so this command-local identity creates no alias.

The selected full identity is never validated against the ordinary declared package name. Production and internal variants retain the production declaration; when production exists, an external variant is admitted only as <production-declared-name>_test. A test-only directory instead establishes one consistent test package declaration itself. The one command-kind rule is that a selected command family declares main/main_test while keeping the finalized ordinary identity unchanged. A source import of that command from another directory rejects as a program before tools; a colocated external command test may reuse the canonical production action. There is no fallback from an empty import path to a declaration name. Relative, absolute, and symlink spellings converge through the canonical directory; two unrelated local directories with the same declaration therefore remain distinct. One bound import path mapping to two directories and one directory acquiring two incompatible ordinary import paths are command-global deterministic errors before any compiler, assembler, archiver, or linker ambiguity. The same check spans variants: an external action cannot hide a different directory's production package behind its derived _test compiler path.

The derivation and diagnostics are implemented symmetrically in cmd/ww/main.c and selfhost/cmd/ww/main.ww. The package coordinator in internal/wwpackage/package.ww supplies canonical selected directories and variant descriptors, preserves an explicitly resolved identity only for one direct request, and forwards a caller's -w semantic-action store unchanged. It never derives identity or persistent layout from a pattern traversal prefix and does not add that prefix to import search. w6c and wcc consume the finalized dotted identity as export/symbol owner while reading the declared name independently from export data; neither tool performs directory lookup or introduces a package registry.

Artifact publication follows the semantic action instead of product order: production uses the full finalized ordinary identity, internal appends -internal-test, external appends _test-external-test, and their generated mains append -main. Generated package identities are likewise derived from the full variant identity, for example __wwtestmain.a.foo.internal.main and __wwtestmain.b.foo_test.external.main. Equivalent products therefore reuse an already-interned directory action and generated-main action and publish through the same persistent-workdir slots regardless of discovery or product order. The narrow raw single-file compatibility path alone retains __root.

A deterministic dependency-first traversal of the complete command union invokes the compiler, assembler, and in-driver deterministic archiver once per interned action. This is compile-time interning, not linker-argument deduplication. Each product is then linked separately from its generated-main or executable root archive and the complete reachable archive closure. The shared plan remains package-test-specific; it is not a generalized scheduler, action schema, cache, or protocol.

Each selected directory retains the ordinary entry-directory-first resolution context from the local package slice: its directory, explicit -I roots in command order, then the toolchain source root. Same and external variants of one directory share that context; unrelated directory roots never acquire lookup precedence from their request order. When multiple contexts reach one canonical production package, the loader verifies that every directory import binding is identical before reusing its compile action. A different binding is a deterministic package-resolution failure for the roots that reach it, rather than a first-root-wins build.

A production action failure blocks exactly the roots that reach it. A root-local compile or link failure does not suppress a successfully built sibling product: the command records completion per product, and the coordinator can run and report successful siblings while attributing each missing product as a build failure. The single union build and completed test products share the coordinator's existing -j process bound; captured output is still emitted only in byte-sorted directory/package order.

Every source-imported dependency is a production variant, so dependency *_test.ww files never enter the graph. Imports that occur only in selected test files add edges only to that internal or external variant. Each compiler unit contains only its action's owned source set, while its invocation receives only the byte-sorted direct dependency .wwi artifacts as separate inputs. Variant compiles receive --test-package, which validates and retains private @test declarations as compiler-only export metadata without synthesizing an entry point. The distinct generated-main action consumes that metadata from its direct variant export and synthesizes the dispatcher with -T.

The generated-main action, rather than the tested variant, owns the implicit direct test-runtime support edge. The command-scoped plan compiles the common support production package once for the complete test request. The command resolves that edge from the selected toolchain source tree, not the user search path; the support package's own imports are also loaded in that toolchain context. Normally its graph qualifier is test, so an explicit source import test coalesces with the same canonical package. When a real user package occupies that identity, the command presents the runtime edge to the compiler under the reserved __wwtest qualifier. This keeps a production package named test available to external tests. Explicit raw single-file ww test FILE fixtures retain the narrow fused -T compatibility path because an anonymous multi-package raw unit is not a canonical directory package; that path still consumes support as a direct export and emits a root .wwi/.a.

The reserved support action and an ordinary source-imported package test may coexist only because the former is explicitly rebound to the compiler-only qualifier __wwtest. This is the sole role-based directory alias and cannot be created by a source import. The separate expanded-vendor-route exception in section 11.16 is canonical source-tree identity, not a role alias. All ordinary production actions, including an external test's colocated production dependency, use canonical directory identity and the global bidirectional import-path checks above. There is no role-based tolerance for duplicate ordinary import identities and no late product-closure ambiguity to resolve.

The selected internal/external variants and a production action reached by their imports or test-runtime closure are the sanctioned graph nodes that may share a physical directory. This also lets a toolchain package's own tests coexist with the production action required by the test runtime. An internal variant already defines those production symbols, so the colocated production archive is omitted from that product's final link while the production action's dependency archives remain in the closure. The external product includes the production archive. Variant-only archives are never linked into another product. All ordinary logical and physical package-identity collision checks remain unchanged; only distinct expanded vendor routes receive the section 11.16 symlink-convergence exception.

Both stage linkers receive the generated-main archive first, followed by the complete reverse-topological reachable package-archive closure, runtime, and explicit -L/-l values through a structured argument vector. No .wwi or special root .o appears in linker argv, and no fixed flattened command buffer can truncate a large closure. WWstage emits joined -Ldir and -lname arguments accepted by its native linker, while Cstage preserves the equivalent split forms. Generated artifact paths are bounds-checked before any unit is opened, so distinct root keys cannot alias by truncation. Recursive discovery groups by physical directory before sorting filenames, and one stable escape of the canonical discovery directory names the persistent command work directory. It contains neither a declared package leaf nor a product ordinal, so equivalent path spellings and reordered products select the same request state. Every *_test.ww package variant is built even when a file only declares helpers and contains no @test, so its package clause and imports are still checked by the shared loader. Every built variant is run, and a successful compiler-owned dispatcher with empty output is reported as [no tests].

Persistent workdirs keep a global driver/compiler/assembler/stamp identity and per-action committed units. When that global identity is stale, the command first removes every old .unit.ww voucher while leaving artifacts recoverable. It can then commit the new identity even if one root fails: successful actions have freshly committed units, whereas failed or no-longer-requested actions cannot be reused. A retry therefore recompiles the failed action without discarding unchanged canonical packages built for independent products.

Warm reuse compares the staged owner-only unit, committed artifacts, and the actual bytes of each direct dependency export. A changed shared dependency is compiled once; its direct importers are reconsidered once; and propagation stops as soon as a regenerated importer export is byte-identical. Stable semantic artifact keys make that behavior independent of which ordinary or test product first discovered the action.

This ownership split follows the pinned Go 1.26.5 implementation cited in section 11.12: its loader reuses canonical cached package objects, its work builder interns actions by mode and package, and its test loader reuses ordinary production while constructing only the required internal, external, and generated-main variants. WW borrows that command-global ownership boundary without adding Go's build cache, import configuration format, or module system.

11.8 Implemented exact package-tool invocation slice

The local package builder now launches the compiler, assembler, and linker as an executable plus an argument vector in both Cstage and WWstage. No package source path, work-directory artifact path, output path, test-support qualifier, or link-closure member is flattened into a shell command. Paths containing spaces therefore retain one argument boundary from the package coordinator through compilation, assembly, and final executable linking.

WW_W6C, WW_W6A, and WW_W6L each name one exact executable path. They are not shell fragments and are not searched through PATH. With no override, Cstage keeps its w6c/w6a/w6l siblings and WWstage keeps its w6c_ww/w6a_ww/w6l_ww siblings. The coordinator preserves these variables when it starts the one command-scoped package build, so the same contract covers ordinary directory builds, same-package tests, external tests, recursive test requests, and persistent-workdir tool identity. A failed overridden compiler or assembler remains attributed to its owning package in both stages. The Cstage linker also sets executable mode with chmod(2) on the exact output path rather than invoking an ambient command.

Source imports still own the graph, each directory is still one production package, compiler actions receive direct dependency exports as individual arguments beside an owner-only .unit.ww, and links receive the complete per-root .a closure. Repository-native coverage wraps all three real stage tools at executable paths containing spaces, records every argument boundary, inspects .unit.ww, .wwi, .a, and root archive placement, runs the published binary, compares repeated Cstage/WWstage artifacts and traces, and removes one direct export at compiler entry to compare package-attributed diagnostics.

Go 1.26.5 keeps the same responsibility boundary: its work executor passes the selected compiler or linker tool and a constructed argument slice to the builder, while package loading and action construction remain separate (cmd/go/internal/work/exec.go). WW adopts that exact-tool boundary without adding a command schema, generalized action graph, scheduler, manifest, cache protocol, or package-manager behavior.

11.9 Implemented package and runtime library-root parity slice

Cstage and WWstage now apply the same two-root contract. A nonempty WW_SRCLIB selects the toolchain package-source search root. It is appended after the entry directory and explicit -I roots for every package context, and compiler-generated test support is resolved from that toolchain root rather than from a user package that happens to have the same name. A nonempty WW_LIB independently selects runtime artifacts; executable and test links use $WW_LIB/libwwrt.a. If that archive is absent, both stages pass their sibling start.o and syscall.o runtime objects instead. Missing and explicitly empty values use the same deterministic fallbacks in both stages.

Bare logical build and run targets search the current directory, explicit -I roots, then one selected library root. That last root is a nonempty WW_SRCLIB, else a nonempty WW_LIB, else an existing <selfdir>/../../lib, else an existing ./lib, else <selfdir>/../lib. Once a target is resolved, its import graph retains the ordinary entry-directory-first context and its separate source-root policy: without WW_SRCLIB, that policy prefers the repository and ./lib source trees before falling back to the runtime root. This preserves Cstage's existing distinction between locating a requested package and selecting the toolchain source tree used by that package's imports.

The variables select local directories only. They do not add manifests, dependency declarations, network lookup, package-manager behavior, or a second graph: source import declarations still create every language edge, resolved directories still intern to canonical package identities, compilers still receive only direct .wwi exports, and final links still receive the complete reachable .a closure. Directory tests continue to build same-package and external-package variants through the same command-scoped package universe.

Repository-native coverage uses distinct source and runtime roots whose paths contain spaces, resolves a bare logical target found only through WW_SRCLIB, places a same-named decoy under WW_LIB, records exact compiler, assembler, and linker arguments, inspects .unit.ww, .wwi, .a, diagnostics, binaries, and runtime results, and compares both stages. It separately exercises the runtime object fallback and explicitly empty variables. The package-test observer copies the toolchain source library to a selected root, marks the compiler-generated test-support source, and proves that both the marker and the selected runtime archive reach real same-package test builds through the coordinator.

The ownership boundary follows the pinned Go 1.26.5 implementation: source headers supply imports in go/build/read.go, the loader canonicalizes and reuses packages in cmd/go/internal/load/pkg.go, tests remain real package variants in cmd/go/internal/load/test.go, and direct compile dependencies are expanded separately for linking in cmd/go/internal/work/action.go.

11.10 Implemented persistent-workdir driver identity slice

The package driver is now an explicit content input to every persistent -w DIR package action. Both stages copy the exact invoking executable to .wwtool.ww, alongside .wwtool.w6c, .wwtool.w6a, and the mode/format stamp. A warm invocation byte-compares all applicable live executables before considering any committed unit reusable. A missing or changed driver copy invalidates every .unit.ww voucher before compilation; old artifacts may remain recoverable, but none can be reused without a freshly committed unit. That slice introduced workdir format revisions 8 for ordinary builds and 9 for tests. Later package-identity slices supersede those revisions; the current formats are recorded in section 11.16.

This closes a real hidden-input boundary. The driver, rather than w6c, owns canonical directory interning, source-derived graph construction, owner-only unit composition, direct-export argument construction, deterministic dependency ordering, single-member package archive serialization, and the artifact commit sequence. Unit equality alone cannot identify changes to those algorithms, and a manually maintained format number can be forgotten. Exact driver bytes conservatively cover them during the transitional plain-file reuse scheme. This may rebuild after an unrelated driver change, but it cannot falsely reuse a package after a relevant one.

The linker remains outside the recorded package identity because persistent reuse never skips a final link: each invocation reconstructs the complete reachable .a closure, selects current runtime inputs and link flags, and runs the selected linker. Thus a linker or runtime change affects the requested binary immediately without forcing unrelated package compilation.

Repository-native coverage runs a three-directory import graph through copied, independently mutable Cstage and WWstage drivers. Exact compiler, assembler, and linker wrappers prove a cold dependency-first build, an unchanged warm compile/assemble skip with a deliberate relink, and full package invalidation after only the invoking driver's bytes change. The test inspects owner-only .unit.ww inputs, separate direct .wwi arguments, .a, identity files, transitive link order, diagnostics before tool execution, published binary bytes, runtime exit, and stage equivalence.

Go 1.26.5 draws the same semantic line with a richer cache: its build action ID binds compiler/assembler tool identities, configuration, selected source content, and direct dependency content IDs in cmd/go/internal/work/exec.go, while its link action ID separately binds linker configuration and the package closure in cmd/go/internal/work/exec.go. WW adopts only the correctness boundary in its existing inspectable cmp-based workdir. It does not add build IDs, hashes, a CAS, an action graph, a scheduler, or a manifest.

11.11 Implemented directory-only source-import slice

Cstage and WWstage now use directory packages for every parsed source import, including imports selected only by a package-test variant and real imports in the test-support package. Source imports first probe the bounded vendor candidates in section 11.16, then make one ordered fallback pass for <root>/<import-path>/; they never probe <root>/<import-path>.ww. The compiler-generated edge to test support remains synthetic. Unit composition consequently writes only the owning package's byte-sorted source files and never copies a dependency interface or imported source body. Missing imports retain the importing source position and the same stable diagnostic in both stages.

Root selection remains a separate compatibility boundary. A literal .ww CLI target, or a bare CLI target found as <root>/<name>.ww after the global directory search, can still create one raw single-file root. Its historical inline package clauses may satisfy compiler-fixture bindings inside that raw unit. Directory roots cannot use that exemption, and no filesystem source import can reach it. This preserves low-level compiler fixtures without weakening package-graph identity.

The self-hosted tools no longer depend on the removed behavior. w6a/ and w6l/ are executable package main directories whose sorted source sets are compiled once. The compiler backend is one wcc/ directory package with a narrow exported check/codegen façade; w6c and wwdump import that package from its parent search root instead of importing its implementation files. Legacy test fixtures were converted to directories, except for one intentional compiler leaf-collision probe that now invokes w6c on an explicitly composed raw unit. The Lisp example likewise imports a lispcore/ directory package.

The focused native regression puts only example/foo.ww in an earlier import root and a two-source example/foo/ package in a later root. Both stages select the directory, emit the exact sorted package-owned unit, consume its direct dependency export, produce byte-identical deterministic .wwi and .a artifacts, link and run a transitive archive closure, and repeat the resolution through a real directory-package test. With only the file root present, both stages reject the import with byte-identical package-attributed stderr. The existing exact-tool observer remains the non-duplicated proof that each canonical production action compiles once, compiler units contain only owned sources, compile argument vectors contain exactly direct .wwi inputs, and linker vectors contain the complete reachable .a closure and no .wwi path.

11.12 Implemented direct compiler-export input slice

Cstage and WWstage now share one small package-compiler convention: --import <canonical-import-path> <export.wwi> may repeat before the one owning source unit. It is valid only with -c; paths must be nonempty, strictly sorted, and unique. Both compilers read and parse every export independently under the supplied canonical identity before parsing the owner unit, then pass the merged semantic declaration list through the existing checker, deterministic export writer, and primary-only code generator. Missing export bytes therefore fail at the compiler boundary as w6c: import <path>: cannot read <file>, followed by the driver's stable owning-package attribution. No import configuration file, manifest, schema, package database, or network lookup is involved.

Both drivers construct those arguments directly from the package node's sorted, deduplicated outgoing edges. They never walk grandchildren for compilation. Every .unit.ww contains the node's byte-sorted source files, reset separators, and any sorted driver-private vendor/import-map voucher comments, but no dependency body. When exact source spelling differs from a selected expanded canonical identity, the drivers also pass sorted, unique --import-map <source-spelling> <canonical-import-path> triples. A map must target an ordinary direct --import; it adds no export input or graph edge. The compilers rewrite only the matching primary import's canonical semantic key before merging interfaces. They then obtain that target's declared package name from its direct export and install it as the default qualifier in the owning source file. Source spelling and position remain intact; generated/synthetic imports use their explicit compiler-owned bindings. Executable linking independently walks the full reachable package closure and passes archives, never interfaces. The same path handles an ordinary package, the production-plus-internal-test variant, the external test package and its reused production package, compiler-generated test main, and the reserved test-support package. Persistent workdirs compare a newly emitted export with its committed predecessor before allowing a direct importer to reuse owner-identical artifacts, retaining correctness without a new cache schema or identity record.

The canonical-root regressions additionally prove the following in both stages: a literal root under one import root publishes its complete dotted identity; logical, literal absolute, equivalent, relative, and symlink routes emit byte-identical .unit.ww, .wwi, and .a; root-only and combined root/import requests emit those same bytes; dependency-first and root-first discovery each compile the shared production action once; recursive a/foo and b/foo directories declaring the same package foo publish distinct a.foo and b.foo variants; and two outside-root package foo directories coexist in one command under distinct reversible local identities. Equivalent recursive spellings reuse the same persistent semantic-action store without new compilation, while source imports of the reserved local namespace reject before tool invocation. The exact-argv command root declares package main but keeps its non-main canonical identity in the unit, export, archive, compiler argv, and linker argv; the bootstrap self-rebuild independently proves the same rule for the real w6a and w6l command directories. A focused command-test row proves distinct production, internal, external, and generated-main actions, colocated external-production reuse, parser-only command markers, stage-equal unit/export/archive bytes, both test binaries' runtime behavior, and pre-tool rejection when another directory tries to import the command.

The exact-argv regression uses the real diamond base -> {left,right} -> root. It proves one compile per node; no input for base; only base.wwi for each middle node; only sorted left.wwi and right.wwi for root; exact owner-only unit bytes; the complete four-package link closure including the root archive; no link-time .wwi; exit status 42; and byte-identical units, exports, archives, executables, and tool argument vectors across two clean Cstage builds and two clean WWstage builds. The existing directory-package variant regression checks separate production, internal, external, and generated-main actions, exact generated-main direct variant/support exports, canonical production reuse across ordinary and test products, and archive-only link closures. It also reverses equivalent product descriptors and compares exact compiler and linker trace bytes, then changes a shared direct export in persistent workdirs to prove propagation through direct importers stops at the first byte-identical regenerated export. Both stages compile and run those actions with owner-only units and byte-identical artifacts.

The pinned official Go 1.26.5 tag (commit c19862e5f8415b4f24b189d065ed739517c548ba) supplies the design boundary:

• go/build represents one selected directory package with its import path, package name, ordinary files, internal-test files, external-test files, and their imports (build.go, lines 436–493). Those fields remain separate in Go; WW's final-component/name equality is its existing language validation layered on the canonical identity, not a claim that Go conflates Name with ImportPath. Its directory reader is required to return name-sorted entries (lines 108–111), and local directory loading reverse-derives a complete import path by checking GOROOT/src first and then GOPATH roots in order. A candidate under a later root is rejected when the same relative path resolves through an earlier root to another directory; an outside-root directory remains without an ordinary import path (lines 612–665). Forward import lookup selects one directory in search order (lines 725–767), and the selected directory alone is scanned (lines 859–913). The sorted scan assigns each accepted source to that package's ordinary, internal-test, or external-test list (lines 948–1036).
• cmd/go/internal/load derives an outside-root local directory's deterministic pseudo-import path from its slash-form absolute directory and establishes the package-data cache/promise boundary around that resolved key (pkg.go, lines 633–647). WW uses the same reserved full-directory principle but a reversible byte escape, strengthening it so two canonical directory spellings cannot collapse merely through character sanitization. The Go loader expands source imports before recording their canonical paths (pkg.go, lines 658–669). It resolves canonical path and directory before consulting the package-data cache (lines 833–842, lines 863–911), and the command-global package cache returns the existing package pointer for a later root or import of the resolved identity (lines 757–775). Go keeps a command-line Name == "main" package as the selected command and rejects an import from a different directory, while permitting the same-directory test-loader edge (lines 799–805). The package's parsed import list becomes its direct package dependencies, rather than a transitive flattening (lines 433–440, lines 2024–2047).
• cmd/go/internal/work keys its action cache by operation mode plus package pointer (action.go, lines 202–206) and returns the already-interned action for that key (lines 437–447). A selected package whose independent Name is main receives a link action (lines 450–455) while retaining its ordinary compiled archive action. A compile action depends on only p.Internal.Imports (lines 628–658); an executable link asks for that same cached root compile action (lines 919–957) and separately expands the complete transitive link closure (lines 1034–1068).
• The work executor derives compiler package mappings from those direct build dependencies (exec.go, lines 864–884), compiles the package's own source list to _pkg_.a (lines 928–935), packs and publishes that package archive (lines 1017–1033), and links the compiled main archive with mappings for every dependency already expanded onto the link action (lines 1592–1624, lines 1635–1647).
• Go's test loader explicitly models production, internal production-plus-test, external _test, and generated main, and states that ptest == p when production can be reused (test.go, lines 85–102). Test imports use the ordinary load cache and compare canonical ImportPath (lines 118–161); the internal copy is created only when needed (lines 175–226), while external and generated-main packages remain distinct (lines 228–293). Generated main receives its direct support and selected-variant imports (lines 307–358), and copy-on-write rewriting preserves the original package pointers/actions for unaffected importers (lines 421–472).
• Unified export production begins from the local package, re-exports required dependency data, and prunes unnecessary detail (unified.go, lines 147–168). It type-checks the package's parsed sources and writes deterministically ordered public/private roots (lines 314–362), then finalizes self-contained export data with sorted relocated declaration and body indexes and a fingerprint (lines 463–570).
• Compiler import handling canonicalizes each source import and rejects self import (import.go, lines 125–167), then independently opens and decodes each direct package archive/export (lines 170–225). The complete unified section and linker fingerprint are read from that selected package file (lines 229–296). ReadPackage reconstructs a package from its public export root (ureader.go, lines 28–62), interns embedded package descriptors by canonical path and restores their import lists (lines 152–196), and reconstructs declarations from relocated export records (lines 391–468).

WW adopts those practical ownership and action semantics while retaining its small direct CLI representation and existing self-contained .wwi encoding.

11.13 Implemented unbounded semantic package-identity storage slice

Canonical package identity is no longer stored in, derived from, or bounded by one internal filesystem component. In both drivers every package action now keeps these values separately:

• path: the complete compiler/import identity;
• import_base: the complete canonical ordinary directory identity;
• canon: the complete canonical directory location;
• variant and role: the semantic test/action tags; and
• storage: an internal scratch basename that is never passed as package identity.

canon participates in command-local directory-action interning, diagnostics, and storage-address derivation. The complete path, together with the semantic variant and role, is the persisted/compiler owner carried through source-import edges, module-reset and export ownership, compiler --import arguments, generated-main construction, and symbol qualification; storage participates in none of those identities. The reversible outside-root form is __wwlocal.p<escaped-canonical-absolute-directory> and is allocated to its exact length. It is neither truncated nor replaced by a digest, and __wwlocal remains unavailable to source imports. Package declarations classify package kind and contribute semantic export content; they do not validate a path leaf, supply a missing identity, or alter a command package's canonical path.

Short actions retain their established .unit.ww, .wwi, .s, .o, and .a basenames when the basename plus .unit.new fits the 255-byte supported filesystem component bound and the complete path fits the host pathname API. An action that does not fit uses this bounded storage locator:

__wwpkg.v<variant>.r<role>.h<lowercase-sha256>

The SHA-256 byte input is exactly:

"ww-package-storage-v2:"
|| ASCII(<variant> ":" <role> ":")
|| complete semantic path
|| NUL
|| complete canonical directory

WW package paths and host paths cannot contain NUL, so that boundary is unambiguous. Variant and role are present in both the digest input and the visible locator tag. The digest is only an action-storage address: units still begin with //ww:module-reset <complete-path> and may append deterministic driver-private resolution comments, while exports begin with //ww:module <complete-path>, compiler imports carry the complete path, and qualified declarations use it in generated symbols. A selected executable entry retains its intentional bare linker spelling. User-selected -o publication paths bypass this derivation completely.

Storage assignment is command-global and finishes before any tool is invoked. If two actions prefer the same legacy basename, every unhashed member is readdressed through the complete-action formula instead of rejecting a valid package graph. If two already-addressed, unequal complete actions ever produce the same locator, both drivers issue the same full-identity storage-collision diagnostic before compilation. A persistent workdir also validates every existing regular .unit.ww voucher against the requested complete semantic owner before stale-tool invalidation or reuse; a missing voucher is cold state, while a malformed, non-regular, or wrong-owner voucher is a pre-tool error. Thus ordinary preferred-name collisions are resolved, and a digest collision cannot silently alias two live or warm package actions. That storage slice introduced build version 10 and test version 11 so an older flat-layout voucher was never accepted as current state. Section 11.16 records the current superseding formats.

The persisted semantic owner is deliberately the complete canonical import path, not the canonical host directory: host location must not enter compiler artifacts. Variant and role are encoded in the locator itself. Canonical directory remains part of command-local action interning and the digest input, so distinct locations normally receive distinct slots. In the hypothetical case that two locations with the same import path, variant, and role also collide in SHA-256, they are still one semantic package identity: the freshly composed owner unit must byte-equal the committed unit before reuse, so different sources rebuild and identical sources produce the same deterministic artifacts. A digest collision between different semantic paths fails the existing complete-path owner check before tools. This preserves collision checking without serializing machine-specific canonical directories into units or exports and without a sidecar, registry, or new metadata protocol.

The package coordinator derives no persistent container from a request directory or pattern. -w DIR names the driver's semantic-action store itself, and the driver independently derives every action locator in that store from the full semantic tuple above and validates each committed semantic owner. Equivalent direct, logical, relative, absolute, dotted, recursive, duplicate, and explicit-root-symlink requests that select the same canonical actions can therefore reuse the same slots; request shape cannot split or alias persistent state. For a delegated request only, a missing caller -w directory is created after graph, identity, visibility, cycle, closure, and output preflight. An explicit root symlink is followed and canonicalized; a source entry symlink to a regular file is followed under the entry name, a source-shaped symlink to a directory is ignored, and symlinked recursive children are not traversed.

The old SEP_IMPORT_PATH_MAX and all corresponding 255-byte WWstage import, variant, local-identity, and generated-main checks are removed. Package names, canonical identities, reverse-resolved dotted identities, generated-main identities, compiler import paths, and Cstage assembler symbols now use exact allocations. Cstage's assembler no longer copies a line, operand, TEXT, or DATA symbol through 256-byte arrays, and the C checker no longer resolves a qualified type through a 128-byte prefix buffer. PATH_MAX remains only at actual host pathname and syscall boundaries; the 255-byte constant remains only as the conservative internal basename component bound. The former SEP_MAXPKG, SEP_MAXPRODUCT, and SEP_MAXCONTEXT action-count limits are removed by the dynamically sized package-universe implementation in the next section.

The existing native observers exercise the new boundary with ordinary dotted identities over 255 bytes and punctuation-heavy reversible local identities over 255 bytes. They inspect the complete owner in units and exports, exact direct sorted/deduplicated .wwi compiler inputs, long mangled assembler symbols, complete archive-only link closures, runtime results, and independent Cstage/WWstage artifact bytes. Two deep outside-root directories declaring the same leaf coexist under distinct reversible identities. A long command package and its production, internal, external, and two generated-main actions remain distinct. Logical and literal roots, dependency-first and root-first discovery, equivalent and symlink spellings, reordered products, and warm requests reuse the same production action and persistent slot. The persistent diamond observer also changes a shared dependency export, proves rebuilding of its direct importers, and proves propagation stops when the regenerated export is byte-identical.

This separation follows the pinned official Go 1.26.5 tag at commit c19862e5f8415b4f24b189d065ed739517c548ba, without adopting Go's cache, build IDs, importcfg, module machinery, or scheduler:

• Source imports are expanded to canonical paths before being recorded (pkg.go, lines 658–669, pkg.go, lines 1150–1178). Local directories receive a deterministic pseudo-import identity derived from the full directory while directory and import path stay separate (pkg.go, lines 633–647, pkg.go, lines 863–907). Resolution/package caches retain and reuse those complete values (pkg.go, lines 833–842, pkg.go, lines 909–985, pkg.go, lines 1008–1029).
• go/build reverse-resolves directory ownership through ordered roots and performs the matching forward lookup while keeping directory, import path, declared name, and source lists separate (build.go, lines 612–665, build.go, lines 725–767, build.go, lines 436–493).
• Go interns an action by operation and package pointer, returns an existing action for repeated compilation requests, and attaches only direct package dependencies (action.go, lines 202–206, action.go, lines 437–447, action.go, lines 628–658).
• Its Action keeps Package, Objdir, and Target as independent fields, then assigns a short bNNN/ object directory unrelated to import identity (action.go, lines 84–109, action.go, lines 383–394). Direct dependency identities map independently to archive paths, and the owner is compiled to fixed _pkg_.a (exec.go, lines 864–884, exec.go, lines 928–935, exec.go, lines 1017–1033).
• Linking reuses the compiled root action and separately expands the complete reachable archive closure (action.go, lines 919–957, action.go, lines 1034–1068, exec.go, lines 1592–1624, exec.go, lines 1635–1647). User -o separately controls the publication target (build.go, lines 508–548).
• Go's test loader models ordinary production, internal production-plus-test, external _test, and generated-main packages separately, reusing ordinary production when possible (test.go, lines 85–102, test.go, lines 175–226, test.go, lines 228–293).

11.14 Implemented dynamically sized command-global package universe

Package, dependency, resolution-context, selected-product, traversal, support, order, and closure storage no longer has an arbitrary 256-element boundary. This is a storage correction, not a new build abstraction: source imports still form one command-global canonical package graph; each semantic package variant still has one action; each compiler still receives exactly its direct exports; and each executable linker still receives its complete reachable archive closure.

The Cstage representation is exact and deliberately small:

• sepgraph.pkg is a dynamically allocated struct seppkg * with logical count n and capacity pkgcap;
• sepgraph.context is a dynamically allocated struct sepcontext * with logical count ncontext and capacity contextcap;
• each seppkg.deps is a dynamically allocated int * with ndeps and depcap;
• each seppkg.context_state is a lazily extended, zero-filled unsigned char * with context_cap;
• each package owns a dynamically grown import-occurrence vector recording kind, source spelling, source file, line, column, and stable dependency action index; its separate dependency vector remains sorted/deduplicated;
• parsed sepproduct values are a dynamically allocated vector, and each product stores its support-action index directly; and
• package-load frames and topological-DFS frames are temporary dynamic vectors, replacing recursion proportional to graph depth.

The WWstage representation is isomorphic. sepgraph.pkg: []seppkg and sepgraph.context: []sepcontext use allocated slice length as capacity and keep separate n/ncontext logical counts. Every seppkg owns a dynamically grown bindings: []sepbind, a dynamically grown deps: []i32 with ndeps, and a lazily zero-extended contextstate: []u8. Products, load frames, and topological frames use typed dynamically allocated slices. internal/wwpackage continues to construct one union command for all selected directory-test groups, but now checks the complete 12 + 6*products + 2*includes builder argument count before allocating or starting the driver.

All graph and product growth starts at capacity 8 and doubles until it covers the requested element count. Cstage clamps before INT_MAX, checks the element count against SIZE_MAX / sizeof(element), and publishes a realloc result only after success. WWstage checks against the same signed 32-bit count boundary, allocates a replacement typed slice, copies the live prefix, and publishes it only after success. Context-state growth copies old bytes and explicitly zeros the new tail. The shared deterministic failures are ww: package graph is too large for an unrepresentable count and ww: out of memory for failed storage; compiler and linker argument-count arithmetic is checked before allocation and before any affected tool invocation. Each driver records allocation/size failure during graph discovery and propagates it as a command-fatal load result, rather than treating it as one product's semantic failure and starting tools for a sibling root. The coordinator uses fallible dynamic storage for discovered paths, source/folder/group/plan vectors, process handles, tool environments, and complete builder/run argument vectors; it reports an oversized product set or allocation failure before the corresponding exec.start and cleans an already-created request temporary tree. Host pathname, filesystem-component, process-argument, and available-memory boundaries remain real host constraints; none is used as a disguised package-count maximum.

Vector growth never changes semantic references. Dependency edges, resolution contexts, selected-product roots and variant roots, generated-main/support edges, load/topological frames, order entries, and closure membership are all stable int/i32 indices. Code reserves a graph slot before taking an element pointer and never carries an element pointer across a graph reserve. Capacity, addresses, request order, product order, output names, and workdir location therefore cannot enter action identity, sorting, diagnostics, storage locators, or artifact bytes. Dependency lists retain byte-sorted insertion and duplicate elimination. The iterative loader retains mark-before-child and post-child command-import validation; the iterative tri-color DFS retains deterministic postorder and the complete live path for cycle diagnostics.

No fixed package, product, context, action, support-map, traversal, order, or closure cardinality remains in either driver. The unrelated SEP_MAXLFLAGS == 32 limit is retained solely for the existing -L/-l command-line interface; it neither indexes nor bounds package actions. Compiler and linker tools already allocate their import/input tables from argc; their genuine remaining process boundary is the host's executable-argument limit.

The native package observers generate rather than commit large fixture trees. The extended long_shared_link_closure_is_complete builds and runs a chain of 300 ordinary directory packages under independent cold Cstage and WWstage work roots. Its command root directly imports all 300 packages and repeats one import, proving an action beyond index 256 compiles, the root receives exactly 300 sorted/deduplicated direct .wwi inputs, every ordinary action receives only its one direct export, every .unit.ww contains only its two byte-sorted owner sources, and the linker receives the root plus all 300 archives exactly once and no .wwi. A second command root imports only p000; it reuses all 300 ordinary actions and its exact linker line still contains the root followed by the full p000 through p299 transitive archive chain and runtime archive, proving that closure construction—not the wide root's direct imports—crosses the old boundary. The observer compares every unit, export, assembly, object, archive, and binary across stages. A second equivalent request against the same persistent store invokes no compiler or assembler. Changing p257's export recompiles exactly p257, direct importer p256, and the wide root, and stops before p255 after p256 regenerates a byte-identical export. Closing the chain at p299 -> p000 produces the complete stage-identical 300-node cycle diagnostic with empty compiler, assembler, and linker traces.

dynamic_package_universe_crosses_former_boundary first selects 309 package-test products through 257 directory spellings in one direct request per stage. Fifty-two real directories each produce ordinary production, internal production-plus-test, external _test, and two generated-main actions; the shared support closure brings that command-global universe to 270 actions. Two hundred five explicit symlink spellings of p000 add distinct valid products and resolution contexts while reusing its one canonical internal action. The observer reverses product and variant order between stages, proves one compile per action, checks all 309 binaries exist, runs boundary products, validates variant-owned units, and compares every action artifact plus representative binaries byte-for-byte.

The same observer then exercises the public ww test <tree>/... coordinator path. The generated tree has 257 real directories: the 52 test-bearing directories above plus 205 production-only directories. The coordinator forms and passes 309 products and 258 contexts (including support) in one driver request. Reusing the direct request's persistent semantic-action store preserves all 270 existing actions and adds exactly 410 production/generated-main actions, for 680 distinct actions. Both stages run every product, produce byte-identical ordered coordinator output and all 680 action artifacts, and perform no compilation on a second equivalent public request. It checks all 104 test-bearing result labels and all 205 no-test result labels, while a builder-boundary observer records exactly one invocation containing all 309 descriptors for each cold and warm public request. Existing focused observers continue to prove dependency-first/root-first canonical reuse, root-only/combined artifact identity, exact reordered-product trace bytes, and request-shape-independent persistent action reuse.

This representation follows the semantic separation and scalable action construction in the pinned official Go 1.26.5 source, identified by VERSION, lines 1–2, at commit c19862e5f8415b4f24b189d065ed739517c548ba:

• Go's loader states that repeated package lookup returns the same pointer (pkg.go, lines 633–636), resolves canonical path and directory before package-data lookup (lines 863–911), and reuses the package cached under the resolved ImportPath (lines 757–768).
• A Go builder has one command-global action cache, while each action's dependencies are a dynamically accumulated slice (action.go, lines 38–45, lines 84–89). The cache key is operation plus canonical package pointer and returns the existing action (lines 202–206, lines 437–447).
• Go constructs an archive compile action and dynamically appends actions only for the package's direct imports (lines 628–659). It separately interns a link action rooted in that cached compile action (lines 919–958) and dynamically expands the complete transitive link closure (lines 1034–1068). Requested package actions are likewise accumulated with append (build.go, lines 519–534, lines 551–558).
• During execution, Go builds one package from its own source list (exec.go, lines 721–790, lines 928–935), maps its direct action dependencies into compiler inputs (lines 864–884), and links the root archive with mappings for the complete link-action closure (lines 1592–1647).
• go/build keeps directory, import identity, declared name, and ordinary, internal-test, and external-test file lists separate (build.go, lines 436–493); ImportDir explicitly processes the named directory (lines 521–525), reads precisely that directory (lines 859–900), and assigns accepted files to the separate package-owned lists (lines 948–1039).
• Go's test loader explicitly returns generated main, internal production-plus- test, and external-test packages, reusing production when valid (test.go, lines 85–102); constructs the internal, external, and generated-main variants separately (lines 175–293); dynamically appends, sorts, and deduplicates generated-main imports (lines 315–376); and uses copy-on-write test variants while preserving unaffected package objects (lines 421–474).

WW adopts those package/action distinctions and scalable dependency accumulation, but not Go's build IDs, module system, importcfg, cache/CAS, preloader, parallel action scheduler, or network behavior. Normal local WW builds and tests remain offline, manifest-free, registry-free, database-free, CAS-free, and network-free.

11.15 Implemented internal-package import visibility

A source import whose complete canonical import path contains a directory component named exactly internal is now contextual. Locate the final such component. Its parent directory is the ownership boundary, and the import is legal only when the importing source package's canonical filesystem directory is that boundary or a descendant at a real path-component boundary. Thus domain.client may import domain.internal.secret, while outsider and domainx.client may not. internalx has no special meaning. For domain.internal.outer.internal.deep, the final internal wins and the owner is domain.internal.outer, not domain.

The complete dotted import identity determines whether the rule applies and how many components comprise the final internal plus its following suffix. WW removes those components from that source edge's resolved lexical target route, canonicalizes the resulting owner directory physically, and compares it with the already canonical physical importer directory. Stripping precedes physical canonicalization because a symlink at or below internal may point to a target with a different name or depth; an action spelling interned by an earlier edge is never reused for this contextual calculation. Cstage obtains directory canonicalization through realpath; WWstage uses its equivalent chdir/getcwd canonicalization. For explicit single-file compatibility roots, Cstage uses realpath and WWstage's component walker uses lstat/readlink plus the same canonical current-directory representation; both preserve component and trailing-directory semantics across symlinks. The equality-or-/-boundary comparison never uses a raw string prefix. This division is intentional: an unrelated physical ancestor literally named internal does not impose visibility on an import whose canonical identity has no such component, while a symlink spelling at or below an import-path internal cannot move the effective owner or importer. Explicit single-file compatibility roots use their canonicalized containing directory as the importer context. Directly selecting an internal directory as a command root remains legal because selection is not a source import.

Visibility is an import-edge property, not part of canonical package or action identity. Both drivers resolve the complete target, intern or reuse its one canonical production action, and then perform the importer-context check before accepting the source binding or dependency edge. Consequently an allowed importer may load and compile the target normally, but that cached action cannot authorize a later forbidden importer. Reversing requested products or visiting the forbidden importer first produces the same result. No declared package name, leaf, output name, product ordinal, storage locator, hash, or test variant participates in the decision. Legal edges therefore retain the existing owner-only byte-sorted unit, exact sorted/deduplicated direct .wwi compiler inputs, deterministic archive, and complete archive-only linker closure.

Cstage represents the rule with the bounded sep_internal_parent_count/sep_internal_import_allowed helpers in cmd/ww/main.c. WWstage has the isomorphic sepinternalparentcount/sepinternalimportallowed helpers in selfhost/cmd/ww/main.ww. Neither adds a package field, fixed-size package table, second action universe, or action-key input. A distinct SEP_LOAD_INTERNAL result propagates through the iterative loader. On rejection both stages emit the importing parser position followed by exactly:

use of internal package <canonical-import-path> not allowed

The command returns immediately from graph loading, before directory-identity finalization, unit composition, workdir owner validation, stale-voucher invalidation, tool-identity recording, compiler, assembler, in-driver archive production, linker, publication, or execution. A cold rejection therefore commits no target/importer artifact or .wwtool.* state. A forbidden warm request cannot use a previously committed target to bypass the check and leaves the already committed target voucher and all tool records byte-unchanged. Allocation or path canonicalization failure remains a deterministic command-fatal pre-tool error; the visibility helper does not change stable package indices or growth behavior.

Every selected source file reaches the same driver scan. The rule therefore applies to ordinary production sources, same-package test sources in the internal production-plus-test variant, external _test sources, and real source imports inside the test-support package. Generated-main-to-variant, generated-main-to-support, and coordinator product wiring remain synthetic edges and are not retroactively treated as source imports. No coordinator change is required: cmd/wwtest only dispatches, and internal/wwpackage/package.ww only discovers and classifies source groups, constructs the union request, and runs the driver; neither resolves a source import or owns its importing position.

This follows only the pinned official Go 1.26.5 source at commit c19862e5f8415b4f24b189d065ed739517c548ba:

• loadImport resolves and reuses the canonical package first, then explicitly checks the rule on every import because the result depends on the importing code, and attaches the importing position (pkg.go, lines 787–791).
• The rule is the tree rooted at the parent of the target's internal directory (pkg.go, lines 1463–1471). A package named directly on the command line is not an import (lines 1498–1502), and the import-path boundary is located before filesystem containment (lines 1505–1515).
• Go's filesystem branch cleans the importer and owner, requires a path-component-aware prefix, and retries with both paths symlink-expanded (pkg.go, lines 1534–1546). Its rejection text is exactly the diagnostic above (lines 1564–1571), and its exact-component search deliberately selects the final internal element as the most restrictive rule (lines 1574–1590).
• Same-package and external test imports both pass through ordinary loadImport with their own source positions before variant construction (test.go, lines 102–161). Generated test-main dependencies are synthesized separately (lines 307–330).
• Go's downstream action cache keys canonical package actions independently and consumes the already validated direct package imports, so visibility does not belong in action identity (action.go, lines 437–455, lines 628–657).

The focused native observer internal_package_import_visibility generates its lexical package/import tree under a temporary physical ancestor also named internal, while deliberate symlink destinations live outside that ancestor, so that unrelated-host-path case is exercised rather than documented only. It builds and runs allowed descendant, nested-final, internalx, and symlinked physical-owner cases; rejects outsider, sibling-prefix, nested-final, and symlink-escape cases at exact source positions with empty tool traces; maps an internal target symlinked to a differently shaped physical path back to its lexical owner's canonical directory; builds an internal package directly; and reverses requested allowed/forbidden product-descriptor order around one reused target. It checks ordinary, internal-test, and external-test source imports, generated-main transitive-export isolation, owner-only source order, direct export inputs, archive-only link closure, runtime results, warm no-op package production, rejection-state preservation, and the primary production chain's unit, export, assembly, object, archive, binary, and exact normalized tool arguments across independent cold Cstage and WWstage work roots. Test variants additionally compare their unit/export/archive bytes and exercise byte-equivalent runtime output in both stages.

11.16 Implemented manifest-free local vendor-directory imports

The local package loader now expands only imports parsed from source. For a source package whose resolved lexical route is <root>/domain/app, an import of lib.math probes these directory packages in order:

<root>/domain/app/vendor/lib/math
<root>/domain/vendor/lib/math
<root>/vendor/lib/math
<ordinary ordered-root lookup for lib/math>

The walk stops at that edge's applicable active source root. It never walks an arbitrary filesystem ancestor and never acquires a boundary from another requested product. Each resolution context carries the current package's lexical route and source-root boundary. An explicitly identified root derives the boundary by removing and round-trip validating exactly its dotted identity components. An unbound literal root chooses the first precedence-valid strict ancestor in its own ordered search roots, or the selected directory itself. An ordinary child records the exact root that selected it; a vendored child inherits the parent's boundary. These values are contextual resolution state, not package/action identity.

Before source scanning, a literal directory whose lexical route is representable below that boundary binds the complete relative dotted identity, including any vendor components. A truly rootless literal keeps its reversible __wwlocal identity and cannot be coalesced with a same-physical vendored action. Thus selecting a vendored directory directly remains legal without letting product order donate its action identity to or from a source import.

A candidate shadows outer and ordinary candidates only when its directory contains an observed non-directory name ending in .ww. This deliberately includes _test.ww and the bare name .ww, matching Go's suffix probe; an actual subdirectory named x.ww and an unreadable candidate with no observed source do not shadow. Once a candidate is selected, normal package enumeration reports its real errors, including a test-only or otherwise production-empty directory, rather than falling through.

Source retains only the effective spelling, such as lib.math. Selection assigns the target the complete expanded canonical identity, such as domain.app.vendor.lib.math, and separately canonicalizes its physical directory. The target's declared name independently supplies the default qualifier in each importing source file. The action key is that expanded identity and canonical directory, plus the existing variant/role. Different physical vendor copies are distinct; different expanded vendor routes remain distinct even when symlinks converge on one physical directory; repeated resolutions of the same pair reuse one action. Source spelling, importer context, product order, output name, declared leaf, and allocation/discovery order do not enter that key.

The final exact non-terminal dotted component named vendor determines the effective suffix and owner. A nested path uses its final vendor; vendorx is ordinary; and a path ending exactly in vendor names an ordinary package. After resolution and action interning, every source edge first performs vendor visibility and then verifies source spelling. An allowed importer that directly spells an expanded path receives the source-position diagnostic:

<expanded-path> must be imported as <effective-suffix>

An outside importer receives, with visibility taking diagnostic precedence:

use of vendored package not allowed

Directly selecting a directory below vendor remains legal because a command root is not a source import.

Visibility derives the owner from the current edge's resolved lexical vendor route and only then canonicalizes that owner physically. It never strips components from the already-canonical target, whose symlink shape may have a different depth. The importer is its canonical physical directory. Equality or a real /-component descendant is allowed; a raw string prefix is not. Consequently an importer reached through a symlink is judged by its physical containment, while a vendor target symlinked to a differently shaped physical directory keeps the owner established by the lexical vendor route. The check runs after intern/reuse on every source edge, so an action loaded by an allowed importer cannot authorize a later forbidden spelling or importer.

Cstage represents the resolution with route and source_root in sepcontext, the bounded sep_resolve_source_import and vendor helpers, stable integer action references in typed source bindings, and transient {package,context} loader children. WWstage uses the isomorphic sepcontext, sepresolvesourceimport, typed bindings, and transient child vector. The package-global dependency set remains the sorted/deduplicated canonical action set used for compilation and linking; contextual child traversal never creates a second universe or contaminates action identity. All new storage grows with checked allocation, and package/context vectors continue to expose only stable integer references across growth.

A compile still receives one sorted, deduplicated --import <expanded-path> <dependency.wwi> triple for each direct dependency. When source spelling differs, the driver additionally supplies the sorted, unique auxiliary mapping:

--import-map <source-spelling> <expanded-path>

Both compilers require the map target to be an existing direct --import, require source keys to be sorted and unique, and require a matching import in the primary source input. No leaf-equality condition exists. After parsing the primary input and before prepending imported interfaces, the compiler replaces its semantic import key while preserving source position and spelling. It then reads the expanded target's declared name from the direct .wwi and installs that name only in the declaring source file. Thus the map adds no dependency or export input, expanded identity flows into self-contained .wwi ownership and symbols, and direct exports still have no transitive leakage. Linking remains independent: each executable consumes its root archive and complete reachable archive closure, never .wwi or import-map inputs.

Persistent units record resolution identity in ignored, deterministic comments:

//ww:vendor-dir <hex-canonical-package-directory>
//ww:import-map <source-spelling> <expanded-path> <hex-canonical-target-directory>

Hex encoding keeps arbitrary legal filesystem bytes inside one comment. The metadata makes ordinary-to-vendor changes and vendor symlink retargeting invalidate the importer even when its source bytes and both already-warm export bytes happen to match. It contains no dependency body and does not alter source positions. The current workdir formats are build 14 and test 13. An equivalent warm request remains a package-production no-op; an export change propagates only through ordinary direct-export comparison.

Production, same-package internal-test, external _test, and real source imports inside the test-support package all use this source-edge resolver. Generated-main-to-variant, generated-main-to-support, and coordinator product edges remain synthetic and receive no map or retroactive source legality. cmd/wwtest remains a dispatcher. internal/wwpackage still discovers and classifies source groups and submits one command-global union; it performs no vendor resolution. It creates only its removable command-temporary coordination tree. The private driver creates a missing delegated -w directory and a requested output directory only after all semantic graph and output preflight succeeds; a rejected request therefore leaves neither directory behind.

Resolution, identity collision checks, contextual legality, dependency-failure propagation, cycles, command kind, publication paths, and action closures all finish before cold scratch acquisition, product status removal, stale-voucher invalidation, tool identity records, producer execution, publication, or runtime execution. A request whose every product is already invalid returns without acquiring scratch; independent valid sibling products retain the existing isolated-build behavior. A forbidden warm request leaves committed vouchers, status files, tool records, and products byte-unchanged.

This is only manifest-free local source-tree behavior. It does not implement Go modules, module vendor mode, go.mod, vendor/modules.txt, importcfg, build IDs, a package database, a CAS, registry access, or network lookup, and it is separate from the future locked vendor store in section 4.6.

The rule follows the pinned official Go 1.26.5 source at commit c19862e5f8415b4f24b189d065ed739517c548ba:

• source-derived imports are expanded and the expanded vendor path becomes the canonical import path (pkg.go, lines 658–668);
• resolution and canonical package-cache reuse occur before the contextual internal/vendor checks, which still run with the importing position on every source edge (pkg.go, lines 722–796);
• vendor lookup walks importer ancestors nearest-first to the applicable root, requires a source-bearing directory, and records the expanded identity (pkg.go, lines 1213–1263);
• the source-bearing probe accepts any non-directory .go suffix and ignores directory-read failure (pkg.go, lines 1418–1429);
• command roots remain legal, while an expanded source spelling must use its effective suffix (pkg.go, lines 1593–1617);
• vendor ownership uses component-aware physical containment and symlink expansion (pkg.go, lines 1620–1667);
• the final exact non-terminal vendor component controls the rule (pkg.go, lines 1670–1688);
• production, internal-test, and external-test imports retain distinct raw source spellings/positions but each resolves canonically, while generated test-main wiring is synthetic (test.go, lines 85–173, test.go, lines 175–266, test.go, lines 272–373, go/build/build.go, lines 415–493); and
• action caching consumes the already-resolved package graph, while output publication remains a later concern (action.go, lines 437–455, action.go, lines 628–659, build.go, lines 470–558).

The native observer vendor_directory_import_resolution generates every source tree temporarily and runs both stages from independent cold work roots. It proves nearest/outer/root/ordinary selection and the active-root boundary; source-bearing versus empty candidates; distinct and reused actions, including same-physical symlink targets; exact spelling/visibility diagnostics and product-order reversal; final-component, terminal-name, component-prefix, and symlink behavior; direct vendored roots; production/internal/external/support source imports and synthetic-main isolation; owner-only sorted units, exact direct exports/import maps, archive-only link closures, runtime output, stage-equal normalized argv/artifacts/binaries, warm no-op production, and cold and warm rejection-state preservation.

11.17 Implemented manifest-free recursive local package-pattern selection

Build and test now share one local request selector. A positional spelling with no ... is one explicit directory root. In a spelling containing ..., each occurrence in a valid UTF-8 spelling has Go's regular-expression wildcard semantics; an invalid UTF-8 pattern matches nothing, and a final /... also matches the directory before that suffix. WW applies the local matcher to its existing manifest-free DIR/... interface as well as ./..., ../..., and absolute spellings. It does not interpret the non-filesystem portion as a module or registry path. Multiple direct roots and patterns may be mixed in one command. For build, -- before the first positional ends option parsing and every following argument is a package selector, including a spelling that begins with -. Once the first positional has already ended flag parsing, a later -- is itself another package argument.

Pattern expansion is request processing only. For each pattern, traversal starts at the directory prefix before the first ... and is bounded to that physical tree. The selector never obtains a traversal or source-root boundary from another product. It produces canonical physical directory roots; only imports parsed from their real source files add dependency edges. The request spelling, wildcard prefix, match membership, output name, request order, and discovery order never become a package import identity or action key and never enter an import search path.

The recursive eligibility rules are:

• Directory entries are read completely and byte-sorted before processing. Every recursively encountered directory whose basename begins . or _, or equals testdata, is pruned with its subtree. An explicit literal selection bypasses these traversal exclusions, so those directories remain legal direct roots.
• A directory owns only its immediate .ww directory entries. Basenames beginning . or _ are ignored. As in go/build, a source symlink whose target is a regular file is read under the symlink entry's byte-sorted name; a source-shaped symlink to a directory is ignored. A subdirectory never donates sources to its parent.
• A directory with at least one production or *_test.ww source is eligible. A recursively encountered source-empty directory is silently skipped. A direct source-empty root is an error. A malformed source-bearing directory is retained as a root and fails during ordinary package-clause or driver loading; malformed files in excluded or source-empty trees do not poison the request.
• The traversal does not prune vendor. Instead a wildcard cannot consume a non-terminal exact path component named vendor. Thus DIR/... may select a code-bearing terminal DIR/vendor but not DIR/vendor/x. DIR/vendor/... explicitly selects that vendor root and its descendants until another non-terminal vendor becomes a barrier. vendorx is ordinary.

The explicit traversal root is opened after following a directory symlink, as in Go. It may therefore name a target outside the lexical spelling, but the target becomes the canonical traversal boundary. Directory symlinks encountered below that root are never followed, so they cannot escape, create cycles, or change selection. A cyclic explicit root is rejected while canonicalizing it. Relative, absolute, dotted, and explicit-root-symlink spellings that reach one physical package collapse to one canonical root. This physical interning is WW's stronger command-global identity rule; it deliberately avoids Go's few GOPATH cases in which different lexical import paths can retain distinct package objects.

Raw requested spellings are byte-sorted first and each is then lexically cleaned before traversal. Matched source paths are canonicalized, sorted by canonical directory and filename, and deduplicated. Products are then byte-sorted by canonical directory and variant. Duplicate patterns, overlapping patterns, and canonical aliases therefore select one root/product and reuse one action. Go itself processes patterns in argv order and suppresses later package objects; WW performs the stronger final canonical sort required by its request-order-independent command universe. Reversing request or product order does not change roots, diagnostics, normalized tool arguments, artifacts, or runtime output.

An unmatched pattern emits:

ww: warning: "PATTERN" matched no packages

PATTERN is quoted with the pinned strconv.Quote rules, including deterministic escapes for quotes, backslashes, controls, non-printing Unicode, and malformed UTF-8 bytes. Warnings are emitted in the sorted request order. With no remaining roots, ordinary build without -o succeeds as an empty build, while test reports ww test: no packages to test. A build with a non-directory -o reports no packages to build; a directory -o reports no main packages to build. A non-directory -o still requires exactly one production root. An existing directory or spelling ending in / receives each selected command under its canonical directory basename; non-main selected roots receive no named output. Two commands with the same destination basename are rejected rather than overwriting one another. All selection, canonicalization, package-clause, duplicate-output, and unusable-request diagnostics precede producer execution. Raw -o and -w spellings and every derived directory-fan-out output, cold scratch name, persistent tool-record name, and package artifact are bounded and validated symmetrically before tools; a raw spelling that fits but whose suffix or command basename does not fit is rejected with the same Cstage/WWstage diagnostic and no filesystem publication. For a delegated multi-root or recursive -S build, -w is required so the assembly outputs have caller-owned persistent destinations instead of vanishing with the coordinator's temporary plan.

Build and test begin with the same eligible canonical directory set. Recursive build removes a source-bearing root that has only test files; an explicitly selected test-only directory remains an unusable build root and fails. Build creates one production product per remaining directory. Test retains test-only directories and, after selection, constructs the already specified isolated production/no-test, internal production-plus-test, external _test, support, and generated-main actions. Pattern expansion does not create those variants and cannot make one variant visible to another.

Vendor selection remains distinct from vendor import resolution. Selecting a directory below vendor, literally or through an explicitly vendor-rooted pattern, keeps its complete canonical local identity; it is never shortened to the suffix after vendor, and its declaration never renames it. A generic recursive pattern does not expose vendored descendants as ordinary short command roots. Independently, an allowed real source import still searches nearest-first below local vendor, creates the expanded identity described in section 11.16, supplies the required --import-map, and performs spelling and visibility checks for that importer even when the canonical action already exists.

The ownership split is exact:

• cmd/ww/main.c and selfhost/cmd/ww/main.ww recognize recursive and multi-root build/test requests symmetrically and delegate them. Their private build route consumes production product descriptors in one command-global package universe. It records library-root completion without linking and links each command product independently. Recursive build forwards -S, -L, and -l through the same private product route; assembly-only products receive completion markers only after their producer pass succeeds. Both stages allocate the same bounded delegation argv and inherit the existing environment directly, so delegation adds no WWstage-only environment-copy allocation or failure point.
• internal/wwpackage/package.ww owns pattern cleaning and matching, bounded directory traversal, source-bearing eligibility, canonical root/product sorting and deduplication, build-versus-test product classification, output coordination, and deterministic reporting. It does not resolve a source import and does not prepend a pattern root to -I.
• cmd/wwtest remains a dispatcher. The Cstage and WWstage drivers own package enumeration, canonical identity, contextual import resolution, graph loading, variants, compilation, archive construction, linking, publication, and persistent reuse.

Consequently the existing action and tool contracts remain unchanged after root selection. Every action unit contains only the owner's byte-sorted source files. A compiler receives exactly the sorted, deduplicated .wwi exports of direct source dependencies and any required vendor import-map binding. Each executable link receives its root archive and complete reachable archive closure, never a .wwi. Canonical duplicate roots reuse the same action; the pattern text, declared qualifier, and product receiving an output do not affect canonical or persistent action ownership.

Selection and package-clause validation finish before the coordinator creates its removable temporary plan, and the coordinator never creates persistent state. Driver graph, identity, visibility, cycle, closure, and output validation finish before a missing delegated work or output directory is created and before scratch, voucher, status, tool-state, publication, or producer mutation. If later setup of another requested directory fails, both drivers remove every empty path prefix created by that setup while preserving all pre-existing caller directories. A cold rejected pattern leaves no partial request state. The same rejection against an existing caller work root leaves its marker and every committed artifact, voucher, and tool record byte-unchanged and invokes no producer. A valid equivalent direct or recursive warm request remains a package-production no-op; changed exports continue to propagate through direct dependencies only.

This slice follows only official Go 1.26.5 source at commit c19862e5f8415b4f24b189d065ed739517c548ba:

• local literal versus wildcard handling, the prefix before the first ..., explicit root-symlink following, recursive exclusions, source-empty omission, and malformed-directory retention are in cmd/go/internal/search/search.go, lines 276–418, and its quoted unmatched-pattern warning is at lines 424–429;
• wildcard syntax, the empty match for trailing /..., and the non-terminal vendor barrier are in cmd/internal/pkgpattern/pkgpattern.go, lines 32–106, including invalid-UTF-8 rejection at lines 75–76, with non-terminal vendor-element replacement implemented at lines 125–137;
• manifest-free request expansion, canonical package loading, first-occurrence deduplication, and pattern membership are in cmd/go/internal/load/pkg.go, lines 2922–2965, while source imports alone recurse through package loading at lines 2024–2047;
• build uses that common matcher, implements single versus directory -o, and omits wildcard-selected test-only roots at cmd/go/internal/work/build.go, lines 459–559 and lines 731–745;
• test uses the same package request set, reports an empty set, and constructs isolated test variants only afterward at cmd/go/internal/test/test.go, lines 684–719 and lines 1133–1226;
• immediate directory ownership and hidden/underscore source-name exclusion are in go/build/build.go, lines 859–914 (including regular-file source symlink following and symlink-to-directory omission at lines 886–900), parse-error retention and production/test classification are in lines 931–1036, and NoGoError eligibility is in lines 1076–1082; and
• child symlinks are skipped by Lstat-based directory walking and entries are traversed in byte order at cmd/go/internal/fsys/walk.go, lines 14–59 and os/dir.go, lines 109–125.

Build subcommand option termination is delegated by the Go command at cmd/go/main.go, lines 312–321, with -- termination implemented by flag/flag.go, lines 1074–1089, and the parse loop stops at the first positional at lines 1153–1176.

Unmatched-pattern quoting uses the pinned strconv.Quote decoder and escape rules in strconv/quote.go, lines 28–123, including the exported Quote entry at lines 117–123, and the IsPrint algorithm at strconv/quote.go, lines 515–559, with its generated tables in strconv/isprint.go, lines 8–733.

No module cutoff, go.mod, module vendor mode, vendor/modules.txt, importcfg, build ID, registry, database, CAS, or network behavior is copied.

The native observer recursive_tree_discovery generates every tree temporarily and runs Cstage and WWstage from independent cold roots. It covers ordinary and test-only roots, hidden/underscore/testdata exclusions and direct exceptions, terminal and explicit vendor patterns, an imported expanded vendor dependency at both package-local and ancestor vendor directories, overlapping/reversed/duplicate patterns independently in each stage, root, child-directory, regular-file, and directory-target source symlinks, canonical aliases, source-empty and malformed directories, build/test variant selection with an asserted common production-root set, middle-position and valid-UTF-8 wildcard edge cases, byte-sorted quoted unmatched diagnostics, multi-command directory output and duplicate-destination rejection, recursive assembly/link-flag forwarding, build -- termination, raw-versus-derived output/scratch/work-path boundaries with zero producer calls, owner-only units, exact direct exports and import maps, archive-only links, normalized stage-equal tool argv and artifacts/binaries/output, warm no-op package production, and cold/warm pre-tool rejection-state preservation. The existing vendor_directory_import_resolution observer supplies the exact compiler --import-map argv proof for imported vendor dependencies, and the existing diamond and long-closure observers independently prove sorted/deduplicated direct .wwi cardinality and archive-only reachable link closures.

11.18 Implemented canonical identity, declared-name, and file-import-scope slice

Directory-package identity and source naming are now independent throughout local build and test. For example:

// canonical import identity: acme.codec
package wire;

is imported with the existing dotted syntax:

package main;
import acme.codec;
export fn main() i32 = { return wire.value(); };

The package action, exports, symbols, archives, dependency edges, persistent storage ownership, and link closure remain owned by acme.codec. Only the source-file binding is named wire. The path leaf codec is not installed as another qualifier, and a sibling source file receives no wire binding unless that file has its own import.

The implemented representation keeps five facts distinct:

1. source import spelling, including its source file, line, and column;
2. contextually expanded canonical import identity;
3. canonical physical directory;
4. the one declared package name read from eligible source clauses and .wwi package markers; and
5. the source-file-local default qualifier binding from that declaration to the canonical target.

Both drivers retain one dynamically allocated sepbind occurrence for every real import site. The occurrence stores the source spelling and position plus a stable action index. Contextual internal and vendor resolution is therefore rechecked for every source import even if its target action already exists. A separate package dependency vector unions those occurrences, deduplicates by canonical action, and sorts by complete canonical identity/variant/role. That vector alone supplies dependency traversal and compiler --import arguments. Patterns, qualifiers, declared names, export closure facts, and generated edges never create source dependency edges.

The owner-only composed unit preserves byte-sorted source boundaries with one //ww:module-reset <canonical-owner> separator per file. Parser nodes carry a source-section ID as well as canonical owner and declared package name. Each real N_USE therefore belongs to one source section. .wwi emission repeats canonical-owner/package markers as needed for contributing source sections and retains imports only with the declarations from the file that owned them. Interfaces remain source-like transitional data, but canonical owner and declared name are no longer collapsed into one token.

w6c and w6c_ww validate every direct export's leading canonical owner against its paired --import path. After all direct exports are parsed, they build canonical-path-to-declared-name metadata from those interfaces, apply any vendor --import-map only to canonical identity, and bind the imported declaration as the default qualifier of each primary source import. The checkers and code generators select bindings by source-section ID and canonical owner. Qualified and retained bare-import lookups mark only that file's binding used. Two files may consequently bind the same name to different canonical packages, while the graph still contains one edge/action for each target.

Within one file, two imports that produce the same default qualifier are a redeclared binding; the later unused binding is also reported. An unused import is reported at its own import position even if a sibling file uses the same qualifier or canonical dependency. A package-scope declaration collides with an equal import binding from any contributing file, matching Go's reconciliation of package and file scopes. Conflicting production package clauses remain a loader-owned deterministic error before producers. Compiler-owned scope or use errors may invoke the compiler, but the driver removes that action's staged .new unit/export/assembly/object/archive files instead of renaming them. No completion/status marker is written, and no publication occurs after a failed compile.

Package kind follows the declaration. package main, not a path component, marks a command. A path ending in main remains importable when it declares a different name. Any ordinary source import of a package declared main is rejected as package <canonical-path> is a program, not an importable package, regardless of its path leaf. The one pinned loader exception is an external test's exact same-directory import of the command production: it is rewired to the forced-library test copy. WW admits only that canonical colocated edge.

Test naming is likewise declaration-based. Production and internal-test variants use the production declared name. When production files exist, external files must declare <production-declared-name>_test; a test-only directory may establish one consistent package name from its test files, as in pinned go/build. External action identity remains the canonical production identity plus the existing external variant suffix where that production exists. Imports found only in internal or external test files belong only to that variant. Support and generated-main actions retain their isolated identities and archive closures. A generated dispatcher privately binds its tested target through --test-target-package <canonical-path> so a command variant declared main does not collide with the dispatcher's own synthesized main; this is compiler-generated wiring, not source alias syntax. Zero-test dispatchers mark their compiler-owned target/support metadata imports consumed.

Vendor expansion changes only canonical identity and physical selection. A source spelling such as lib.codec can resolve to domain.app.vendor.lib.codec, while the vendored package's declaration, for example package wire, supplies the file-local qualifier. The driver emits one sorted semantic --import-map lib.codec domain.app.vendor.lib.codec and one direct export input despite repeated import occurrences in separate files. The expanded identity continues to own symbols, .wwi, archive, voucher, and link inputs.

Canonical action interning remains the directory/path/variant model of sections 11.7 and 11.14. Independent file bindings never clone an action, and a declared name never enters an artifact basename or storage locator. The name is semantic content in the owner unit and export. Changing only a dependency's declaration therefore keeps the same action identity but changes its export bytes, causes each direct importer to be reconsidered, and stops propagation after an importer's regenerated export is unchanged. An identical warm request remains a producer no-op. Build workdir format 14 and test format 13 prevent reuse of older vouchers that lack these semantics.

Compiler argv still contains exactly the sorted, deduplicated .wwi exports of direct canonical dependencies; no transitive .wwi and no qualifier-derived path appears. Linker argv still contains only the executable root archive and complete reachable archive closure plus runtime/native inputs. Publication and completion occur only after the corresponding action/product succeeds.

Responsibility is intentionally split as follows:

• internal/wwpackage classifies production, internal, and external test files from their declarations, derives the allowed external name from the production declaration, expands request patterns, and submits variant roots. It does not resolve imports, choose qualifiers, or create dependency edges.
• cmd/ww/main.c and selfhost/cmd/ww/main.ww own per-site parsing and contextual resolution, canonical directory/action interning, declared-name consistency, imported-command rejection, sorted dependency union, exact tool argv, variant/generated-main construction, persistence, linking, and publication. Their storage, diagnostics, call positions, and allocation failures are isomorphic.
• cmd/w6c/cmd/wcc and selfhost/cmd/w6c/selfhost/cmd/wcc own export owner/name reading and writing, file-local binding installation, collision and unused-import diagnostics, name/type lookup, canonical symbol ownership, and generated-dispatcher private qualification. Cstage and WWstage emit byte-identical applicable interfaces, assembly, archives, and binaries.
• cmd/wwtest remains only the test-command dispatcher.

The behavior follows pinned official Go 1.26.5 source at commit c19862e5f8415b4f24b189d065ed739517c548ba:

• go/build.Package stores Dir, Name, ImportPath, production files, and test files independently, while package-clause consistency is checked during file classification (go/build/build.go, lines 436–493, go/build/build.go, lines 939–1049).
• The loader interns by resolved ImportPath, keeps ImportPath and Name separate, constructs edges from source imports, and rejects an imported package whose Name == "main" except for the exact same-directory test case (cmd/go/internal/load/pkg.go, lines 633–636, lines 757–806, lines 2024–2047).
• The test loader builds distinct production/internal/external/generated-main packages, derives external Name from p.Name + "_test", handles the same-directory command self-import, and rewrites it to the library-form test copy (cmd/go/internal/load/test.go, lines 144–203, lines 228–293, lines 421–472).
• go/types.Package stores independent path and name. The resolver creates one child scope per source file, uses the imported package's declared name when no alias is present, inserts imports into that file scope, reconciles them with package declarations, and reports unused imports (go/types/package.go, lines 26–40, go/types/resolver.go, lines 237–350, lines 463–480, lines 701–735). The production compiler mirrors those rules (cmd/compile/internal/types2/resolver.go, lines 223–335, lines 473–486, lines 706–740).
• Compiler export import reconstructs and interns package descriptors by canonical path while restoring their independently encoded package names and imports (cmd/compile/internal/importer/ureader.go, lines 152–196, go/internal/gcimporter/ureader.go, lines 224–244).
• Work actions consume cached canonical package objects and direct imports, while test execution builds and links the isolated test action graph (cmd/go/internal/work/action.go, lines 437–455, lines 628–659, cmd/go/internal/test/test.go, lines 1133–1226).

The deliberately retained source grammar is import dotted.path;. Quoted Go imports, explicit aliases, dot imports, and blank imports are not implemented. Because bindings already retain source ownership and position separately from canonical identity and declared name, adding explicit aliases later will not require another package-wide namespace redesign.

The focused native observer declared_name_identity_and_file_import_scope generates every tree temporarily and exercises both stages from independent cold roots. It proves the identity/name/qualifier split, file-local collision and unused behavior, one-edge/action reuse, command and imported-command rules, all test variants, vendor expansion, recursive/direct selection, owner-only units, exact direct exports and archive-only links, warm no-op behavior, declared-name invalidation, rejection-state preservation, normalized argv, artifact/binary identity, allocation-bearing runtime behavior, and request/product-order independence.

12. Candidate architectures and hard-gate decision

Five candidates were developed as coherent systems, not as feature bins.

12.1 Candidate A: Go-like integrated language command

One ww command would infer directory packages/imports, compile/cache/test them, and add a small module/lock layer. Native inputs would remain compiler flags or toolchain conventions. This preserves the strongest part of Go: explicit imports, fast direct export data, and a short ordinary command (Go command design). A WW-specific lock and no-network build could improve on modern Go module behavior.

It still fails as an end-to-end native design. A source-only graph cannot name host generators, C header trees, archive order, linker scripts, CRT, SDK, or sysroot. Ambient compiler-driver and pkg-config behavior would remain outside the key, and the build/host/target triad would be incomplete. Adding typed native actions, complete toolchains, and content records turns it into Candidate E.

12.2 Candidate B: Hare/Odin-style local source plus an outer build tool

WW would use search roots and directory modules, with source vendored or supplied by an OS package manager; Make-like orchestration would own native work. This is small locally and avoids a language-owned network resolver. Hare 0.26.0 (released 2026-02-13) is a useful reference: directory modules, direct textual export data, HAREPATH, explicit system-library flags, and cross-architecture tooling are documented in its official manuals (modules, project structure, system libraries, cross compilation). Odin's named collections are a related local-source convention (Odin overview).

As a complete WW system it fails: ordered search roots are selection policy without locked source identity; ordinary outer recipes expose ambient tools, environment, and mtimes; language and native graphs/caches cannot jointly explain invalidation; and cross sysroots/ABI providers remain project conventions. Making the outer tool hermetic and content-addressed yields Candidate C, not this model.

12.3 Candidate C: two-layer Plan 9-style builder and orchestrator

A strict package builder would compile an import graph. A separate small declarative DAG tool would own generators, C/assembly, images, archives, and links. With content records, pinned tools, sandboxing, and an exact handoff this can pass every hard gate. It reflects mk's valuable complete-graph/parallel-tool shape without copying its mtime and ambient-environment assumptions.

It loses after the gates because the boundary creates two graph protocols, two selection UIs, two explanation namespaces, and either duplicate scheduling/cache logic or a coarse “build all packages” action. Ordinary native projects must know when to invoke each layer. If both front ends lower into one shared scheduler and cache, and ww owns the ordinary invocation, the result is the smaller Candidate E. Keeping the second production tool after that offers no remaining orthogonal concept.

12.4 Candidate D: Zig/Cargo-like programmable integrated project

A manifest would define artifacts and dependencies while a host-executed program constructs a flexible native graph. This handles more native cases than a language-only command. Zig 0.16.0, pinned here to its 2026-04-13 release metadata, provides explicit target/native concepts, hashed package sources, local dependency overrides, and generated-file graph edges (download metadata, 0.16.0 reference, build system). Cargo 1.97.1, shipped with Rust 1.97.1 on 2026-07-16, provides exact package IDs, workspaces, lock checksums, resolver rules, and native links collision handling (Cargo reference, resolver).

Zig's build program and Cargo's build.rs execute to decide or report build behavior; Cargo explicitly documents build-script inputs/outputs and its fingerprint cache (build scripts, build cache). Those reference systems as shipped do not meet WW's complete native/toolchain hard gates.

The strongest coherent D is not left as a straw man: it content-identifies the graph program and host compiler/runtime, declares its whole readable source/tool closure, runs it in the denied-by-default sandbox with no network, and requires it to emit a closed typed graph before artifact execution. That hardened model can pass every gate. It still loses afterward: WW must permanently ship/secure/ bootstrap an evaluator API and host build-program toolchain, users debug both program execution and its emitted graph, dependencies expose framework APIs, and routine exceptions accumulate as library features. Finite records buy the same WW requirements with less user and implementation machinery.

12.5 Candidate E: hermetic integrated action build — selected

Candidate E retains the Go-like ordinary UX and import-derived language graph, then adds only the native/action facts that imports cannot express. Both lower to one typed graph and content cache. It borrows declared tools/inputs and action results from Bazel's hermetic/remote-execution model, and transparent content-derived build records from Nix derivations, without adopting either framework, evaluator, daemon topology, or user interface (Bazel hermeticity, Bazel remote caching, remote execution protocol, Nix derivations).

It passes every hard gate and is selected. Its concepts are exactly package, module/source selection, product/action, toolchain/target, artifact/digest, and native provider/link plan. There is one graph, one scheduler, one cache key, one explanation path, and one ordinary command.

12.6 Hard-gate matrix

Legend: pass means the strongest coherent form has a credible end-to-end invariant; fail means it does not. D denotes the hardened evaluator above, not unmodified Zig/Cargo behavior.

Hard gate	A: Go-like	B: local + outer	C: two layer	D: programmable	E: action build
complete inspectable dependency DAG	fail for native work	fail across tools	pass	pass after sandboxed evaluation	pass
loud cycles/identity collisions	pass for packages; native incomplete	search shadowing fails identity	pass	pass	pass
frozen offline source closure	pass with proposed lock	vendoring can pass	pass	pass	pass
correct cache invalidation	fail for native/tool defaults	fail across mtime/outer recipes	pass	pass	pass
compiler/linker/runtime/toolchain identity	requires the E native layer	fail	pass	pass	pass
explicit B/H/T	fail	fail	pass	pass	pass
correct native dependency/link modeling	fail	fail	pass	pass	pass
deterministic package selection	pass	fail under search roots	pass	pass with lock	pass
reproducible artifacts/boundary	fail for native closure	fail	pass	pass with evaluator certification	pass
bounded bootstrap	pass with a seed	pass with enumerated seed/tools	pass	pass, larger evaluator TCB	pass
explain cache miss/rebuild	fail	fail	pass across two namespaces	pass across evaluation + graph	pass
one simple ordinary path	pass	two commands/configurations	two production layers	pass by convention	pass
no build/test network or mutation	pass if changed from Go defaults	outer recipes cannot guarantee	pass	pass	pass

C, hardened D, and E survive the gates. E wins the requested post-gate comparison:

Criterion	C: two layer	D: hardened evaluator	E: selected
conceptual simplicity	two graph/handoff models	evaluator language/API plus emitted graph	one graph protocol; two finite descriptions
daily usability	user chooses builder/orchestrator	ordinary defaults can hide evaluator, exceptions cannot	ww build/test/run always
clean speed	parallel orchestration can match	evaluation overhead, then parallel graph	direct complete template plus parallel actions
incremental speed	cross-tool handoff may be coarse	evaluator must rerun/cache its own dependencies	direct exports, lazy content keys, cached link
rebuild blast radius	good only with API digests across handoff	can be good after evaluation	public-output digest propagation is intrinsic
cross-compilation	orchestrator can model it	rich API can model it	roles are built into every typed record
FFI/native honesty	outer layer owns it separately	API can express it	same provider/link records as package objects
implementation maintenance	two adapters/engines or coarse boundary	evaluator, SDK compatibility, sandbox, graph engine	one scheduler/store/adapter set
supply-chain security	two policy surfaces	dependency host code remains an evaluator input	one lock/source/action trust model
ecosystem scalability	two tool ecosystems	framework/API accumulation pressure	closed schemas version only for demonstrated facts
debugging/observability	two explanation namespaces	debug program, evaluator cache, then graph	one causal graph/record diff
bootstrap longevity	small outer interpreter possible	compiler/runtime/evaluator enter seed chain	fixed seed plan; no production evaluator

12.7 Why each subsystem is in or out of the core

Package parsing/resolution and export data belong in the core because only the compiler can state the true language graph and public ABI. Module selection and lock verification belong beside it because an import identity must resolve to one deterministic source before compilation. The typed action engine, target descriptors, native providers, and toolchain closure belong in the core because they share invalidation and link correctness with package objects. Test/doc/ install are thin product selections/materializations over that same graph.

Network transport remains a separate operation, though exposed by ww, because fetching is not building. Credential policy, OS package installation, registry hosting, signing authority, remote execution, deployment, and general release automation stay outside. The finite generate action is the boundary: it lets outer domain tools transform declared artifacts without turning WW into their framework.

13. Migration plan

Migration cost does not affect the decision. It is nevertheless material: the current directly implicated scaffolding is at least 7,969 lines across the Makefile, two drivers, two interface writers, package coordinator, and test wrapper, plus 4,728 lines in the focused separate-build/package/byte-identity/ driver tests counted for this audit. Compiler export/import logic, new native adapters, and bootstrap work add new scope not represented by those deletion counts.

The planning estimate is 15–22 engineer-months for the first production Linux/amd64 toolchain, including tests, migration, and deletion, plus 1–2 engineer-months per materially different additional official target/sysroot. This is an estimate for staffing and sequencing, not a reason to retain a weaker architecture.

Every phase below ends in a bisect-clean commit. Experimental components are not installed as a second user-facing build path. Until the point of no return, the old command remains the only production path; after it, the new command is the only path.

Phase 0 — conformance corpus and protocol freeze

• Land the compact checked-in protocol/schema/ modules for WWAR framing, record/enum/union/default/kind assignments, wrappers, and finite digest byte formulas. Freeze the exact schema-file digests. Preserve representation-only preimage records without embedding the algorithms that construct them.
• Generate data-only codec tables deterministically. Keep one schema-aware reference codec and a second independent oracle limited to raw WWAR framing, hashing, and record identity; neither may implement future compiler or build behavior.
• Gate: strict duplicate-key UTF-8/NFC schema loading, byte-identical generation in separate fresh directories, compact valid/invalid vectors, all assignment coverage, stable malformed-length precedence, domain separation, action-key vectors, and record-kind substitution rejection. No production behavior changes beyond adding this conformance gate.
• Package/API, graph/cache, native B/H/T, and bootstrap measurements remain useful fixtures, but become tests in Phases 1, 2, 4, and 6 respectively. They are not Phase 0 semantic answer tables.

Phase 1 — compiler export and package protocol

• Implement deterministic .wwe/.wwlm writing and direct-interface reading in Cstage and self-hosted compiler paths behind test-only entry points.
• Add package identity/alias syntax, strict directory enumeration, compiler import extraction, target suffix selection, cycle/collision/internal checks, and protocol compatibility diagnostics.
• Construct and test the deep public type closure and public type/ABI digest preimages in compiler code with existing standard-library graphs. Reject .wwi input in the experimental path; do not translate it.
• Gate: Cstage/WWstage emit byte-identical vectors, importers open only direct export files, and API propagation stops on unchanged middle exports.

Phase 2 — one action engine and local CAS

• Integrate the frozen WWAR codec into production code and implement typed pure action-record/key functions, the action graph, lazy keys, scheduler, atomic CAS/results, project index, corruption quarantine, graph JSON, environment and sandbox policy, deterministic failure behavior, and causal explanation.
• Add deterministic package/archive/link adapters using the existing compiler, assembler, and linker as explicitly hashed tools. This is a temporary adapter, not a compatibility promise.
• Keep the engine under an internal test binary; the installed ww still follows the old production path.
• Gate: all action-input mutation and failure-injection tests pass; identical warm builds execute no compiler, assembler, archiver, or linker.

Phase 3 — module, lock, source, and workspace layer

• Implement the manifest, lock, work, and vendor text parsers as ordinary typed parser code, then implement the closed grammar, monotonic selector, canonical lock, HTTPS source-index/archive protocol, immutable source store, signatures, explicit add/update/lock/fetch, overlays, vendor index, and canonical source-tree construction. The schema fixes only the resulting record bytes and source-tree digest formula.
• Build/test/doc/install remain network-denied from their first experimental use.
• Gate: frozen offline builds work from project source + complete locked source closure + lock + installed named toolchain; collision/downgrade/hash/ path-normalization attacks fail loudly.

Phase 4 — native, target, and external toolchain closure

• Implement full target descriptors, B/H/T lowering, C/assembly/native-provider records, recursive provider selection/expansion, generated actions/sandbox, object sidecars, exact link-plan construction, sysroot, libc/CRT/loader/SDK/runtime, shared-library installation, and freestanding products.
• Package supported assembler/linker/archive/C tools as immutable external closures. Implement an external-assembler-compatible textual emission backend, then have the pinned assembler produce package.o inside ww.package; port all WW/Plan-9-dialect runtime and user assembly to a declared supported external dialect, and pass object/link parity. An argv adapter alone cannot consume current w6a syntax. Stop relying on WW-owned w6a/w6l before the experimental gate passes.
• Gate: native conflict, archive group/order, linker-script include, host leak, freestanding, shared loader, and at least one real cross-target suite pass.

Phase 5 — repository and consumer conversion rehearsal

• Define the final identity/layout conversion in a one-shot checker/rewriter under tools/migrate-build/: folded files become directory packages; aliases, manifests, native providers, products, generated inputs, and locks are emitted. It is not an import resolver or runtime compatibility layer.
• Until cutover, CI applies that tool to a fresh ignored shadow tree. Tracked production source stays in old syntax, so the old command remains its only production path; the shadow is regenerated, never a second maintained source tree or shipped interface.
• Dogfood the internal engine on the converted shadow of the standard library, compiler tools, tests, examples, install layout, CI/package inputs, and sample downstream consumers. Compare semantics, diagnostics, performance, artifacts where protocols permit, and complete graph explanations.
• Gate: the regenerated full shadow, distribution/frozen-offline build, self-contained installed toolchain outside the build tree, downstream samples, and bootstrap inputs pass without an old-format edge.

Phase 6 — new bootstrap and reproducible release

• Implement/gate the portable C recovery backend and snapshot generator, then generate/check in bootstrap/ww0.c and the fixed plan. Produce stages 1/2/3/4, rebuild stages with the executable engine, compare the actual bytes of every declared output directly, and produce fixed-point and diverse-seed-compilation reports, signed toolchain bundles, and recovery documentation on a clean machine with no WW compiler.
• Gate: stage 2 equals stage 3 and stage 3 equals stage 4 by explicit byte-for-byte comparison, including raw action/result records, in two roots/concurrency levels; project plus complete locked source closure and published named tool closure reproduce every release artifact.

Phase 7 — point of no return and deletion

In one atomic, bisect-clean cutover commit:

1. install the new engine as ww and make it the sole build/test/bootstrap path;
2. switch repository imports, manifests, locks, toolchains, CI, installation, and release jobs to their final forms;
3. delete both old drivers, both .wwi writers, driver-side .wwi concatenation, composed-unit/module-wrapper logic, generic C/self-hosted lexer/parser/checker/ symbol/codegen support for module directives, -w, old import search/file folding, separate test coordinator, duplicated production Make dependency graph, current Cstage bootstrap, and retired owned assembler/linker path; and
4. delete the migration rewriter after all supported consumers have used its released standalone copy; keep only a format-error guide.

The commit does not accept old .wwi, dotted/file imports, -I, -w, raw library searches, or old work directories. No alias, warning period inside the compiler, environment switch, or fallback subprocess retains a dual system.

Phase 8 — consolidation

• Remove temporary parity fixtures that test deleted artifact bytes while retaining semantic, action-key, reproducibility, and bootstrap regression tests.
• Publish migration statistics and archive the old documentation as historical release material outside the live manual.
• Gate: repository search and executable tracing show one graph constructor, one package resolver, one cache, one test route, and one bootstrap route.

14. Validation plan and release gates

14.1 Unit and format tests

• Phase 0 has golden and adversarial vectors only for WWAR representation, source-tree digest bytes, action-key/record identity formulas, typed record assignments, wrappers, and canonical schema JSON.
• Phase 0 tests UTF-8/NFC, duplicate schema keys, unknown wire type/schema field, oversized declarations, truncation, exact-length mismatch, union shape, and record-kind substitution. Case-fold collision, traversal, symlink/device, cache collision/corruption, and semantic record tests land with their owning executable phases.
• Phases 1–4 add behavior tests for CAS tree/result objects, .wwe, .wwlm, manifest/lock/work/vendor parsing, and target/toolchain/native processing; these are not encoded as Phase 0 vector outcomes.
• Resolver vectors for minimum selection, incompatible-major identities, workspace identity preservation, source origin independence, vendor matching, internal packages, aliases, nested module/root-versus-parent-subpackage identity collisions (including /vN), cycles, and target source specificity.

14.2 Package/interface tests

• One- and multi-file directory membership, file-scoped import use, same/external tests, test-only packages, examples/docs, and generated fragments with import/ package rejection.
• Direct dependency interface-open counts equal package-graph indegree, never transitive closure size.
• Private dependency edits preserve importer keys; exported but unused additions rebuild direct importers; unchanged middle .wwe stops propagation; layout, calling-convention, runtime ABI, compiler, target, and profile changes rebuild the exact affected cone.
• Parameter renames, comments, private declaration ordering, and absolute source paths do not change .wwe; semantic/ABI changes do.

14.3 Native and link tests

• C scalar/aggregate/variadic/callback/TLS/unwind ABI probes against an independently compiled C harness for every official target.
• Binding generation changes on header tree, macro map, preprocessor, tool, C ABI, and target; undeclared include access is denied.
• Assembly dialect/CPU mismatch, wrong-format objects, PIC/shared rules, archive extraction order, repeated libraries, groups, whole archive, weak/strong symbols, version scripts, linker-script includes, and deterministic archives.
• Exactly-one libc/runtime/provider enforcement; identical coalescing; duplicate native ABI/symbol conflict; shared SONAME/loader/runtime installation closure; freestanding entry/script with proof that no libc/CRT/loader appears.

14.4 Cross and sandbox tests

• Matrix with B != H, compiler product H != T, and all three distinct where infrastructure permits. A B generator emits an H input; an H binary is never executed during build; T objects never enter the H link.
• Poison host PATH, includes, libraries, SDK, locale, time, home, current directory, and environment. Every attempted undeclared read/write/network/ process/tool access fails with its action identity.
• Cross tests build without a runner, run only with an exact declared runner, and fail under --require-run when none exists.

14.5 Cache and failure injection

• Change each field in section 6.2 individually and require a key change; change each expressly non-semantic observation and require no key change.
• Bit-flip blobs, trees, results, action mappings, tools, export data, objects, and partial files at every publication boundary. Require quarantine/rebuild, never acceptance or broad deletion.
• Concurrent identical publishers, killed compiler/linker, disk full, rename failure, read-only output, interrupted materialization, stale project index, malicious remote mapping, bad cache signature, and remote outage.
• Clean, local-hit, explicit remote-import, and no-cache builds must yield the same result digests. Test executions still run.

14.6 Reproducibility and bootstrap

• Compare every artifact/action record across two absolute checkouts, source/ cache/output roots, usernames, locales, time zones, umasks, concurrency levels, filesystem enumeration orders, and cold/warm caches.
• Verify debug/release, static/shared, hosted/freestanding, generated/native, and signed/unsigned products. Impure profiles must state exactly why they are outside the byte promise and must never enter shared cache.
• Build ww0 with each supported host C toolchain, reach stages 2/3/4, run the semantic then raw-record fixed point and diverse seed compilation, corrupt each stage input, and recover on a host with no WW installation.

14.7 Performance and migration gates

On the audit's fixed eight-CPU reference host, the first release MUST:

• run package compilation in parallel and complete the full clean toolchain build no slower than the measured 15.515 s make -j8 baseline;
• perform a warm 15-package build with no compiler, assembler, archiver, or linker process and no slower than the measured 0.053 s driver baseline;
• read only direct exports and avoid composed-unit duplication;
• store one CAS copy of duplicate tool/package content across all products;
• produce fully path-independent official artifacts, including host-side tools; and
• provide a typed explanation for every deliberately induced rebuild.

Before cutover, every tracked current package/test/install/bootstrap consumer has an assigned new identity and a passing converted test. The cutover gate includes a repository-wide search for old forms and executable traces proving no old driver, interface, workdir, library search, or test-coordinator path executes.

15. Evidence appendix

15.1 Research method and version pins

Research used official documentation, standards/manuals, release metadata, and current upstream source—not comparison articles or community summaries. The evolving-system snapshot was taken 2026-08-09:

System	Pinned snapshot used
Plan 9	live official 9p.io Volume 2 documents, accessed 2026-08-09; pages are not versioned/dated editions
Go	online docs accessed 2026-08-09; Go 1.26.5 go1.26.5 source tag
Hare	online docs accessed 2026-08-09; Hare 0.26.0 source, released 2026-02-13
Odin	online docs accessed 2026-08-09; dev-2026-07a, commit 819fdc7a80667498b8b365999f1475a66c358640
Zig	Zig 0.16.0, official metadata release date 2026-04-13; source archive SHA-256 43186959edc87d5c7a1be7b7d2a25efffd22ce5807c7af99067f86f99641bfdf
Rust/Cargo	Rust/Cargo 1.97.1, released 2026-07-16; Cargo 0.98.0 commit c980f4866141969fab6254a680546a277789d6f0
Bazel	Bazel 9.2.0 documentation/source
Nix	Nix 2.35.2 manual/source

Zig's separate bootstrap-source archive inspected for this decision had SHA-256 2a8266a4205772ef40838c8cbdf14875855a515ff3adf89b49c2d2ae93613d10. These pins matter because programmable-build and package behavior changes between releases; this document does not generalize an old Zig/Cargo observation to an unidentified current version.

15.2 Primary-source findings

Pike, Plan 9, and early Go. Pike's sources support explicit computable imports, cycle rejection, direct compiled export information, fast compilation, and orthogonal concepts. The collective Plan 9 papers add system-wide placement of complexity and transparent encodings. Plan 9's namespace papers demonstrate contextual filesystem composition; this document infers that contextual location must not serve as WW's versioned distribution identity. mk shows complete graph scheduling while still relying on timestamps, recipes, and environment. The architecture borrows the former principles and replaces the latter ambient assumptions.

• The Go Programming Language, 2009
• Go at Google: Language Design in the Service of Software Engineering, 2012
• Simplicity is Complicated, 2015
• Go in Go, 2015
• Plan 9 overview
• The Use of Name Spaces in Plan 9
• Maintaining Files on Plan 9 with Mk
• Plan 9 Mkfiles
• Plan 9 compiler suite

Later Go. The original Go command demonstrates source-derived package DAGs, directory conventions, and compiler-owned dependency work. The modern module reference documents module identity, Minimal Version Selection, major-version paths, checksums, and commands that may resolve/download modules. Current command and source-install documents also separate build cache/toolchain/bootstrap behavior. These are evidence, not automatic WW defaults; in particular WW uses an exact lock and forbids implicit build-time acquisition.

• About the Go command
• Go module reference
• go command reference
• Go toolchain selection
• Installing Go from source
• Perfectly Reproducible, Verified Go Toolchains
• Go's supply-chain security
• go1.26.5 source

Local-source systems. Hare demonstrates how far a disciplined directory module/search-root system can go with little package machinery; its documented system-library and cross interfaces also expose why raw host paths and tool defaults are insufficient for WW's hard gates. Odin's collections reinforce the local namespace option but do not add a locked whole native closure.

• Hare modules
• Hare project structure
• Hare system libraries
• Hare cross compilation
• Hare 0.26.0 source
• Odin overview
• Odin pinned source

Integrated project systems. Zig supplies useful target/native vocabulary, source hashes, local aliases, declared generated-file edges, and cross-building. Cargo supplies package IDs, exact lock checksums, workspace behavior, resolver documentation, and a native links uniqueness rule. Their programmable build program/script model is deliberately rejected; a dependency host program is a larger and less inspectable abstraction than WW's finite action record.

• Zig build system
• Zig overview
• Zig 0.16.0 language reference
• Zig release metadata
• Zig 0.16.0 source
• Cargo 1.97.1 reference
• Cargo current reference entry
• Cargo 1.97.1 resolver
• Cargo current resolver entry
• Cargo 1.97.1 build scripts
• Cargo current build-script entry
• Cargo 1.97.1 build cache
• Cargo current build-cache entry
• Cargo pinned source

Hermetic invariants. Bazel defines useful distinctions between declared actions, execution platforms, action results, and content-addressed remote data. Nix derivations demonstrate an inspectable build record whose output depends on declared inputs. WW borrows those invariants only. It rejects Bazel's general rule/ecosystem machinery and Nix's evaluator/store-as-package-manager as excess for one language toolchain.

• Bazel current hermeticity entry
• Bazel 9.2.0 hermeticity
• Bazel remote caching
• Remote Execution API
• Bazel 9.2.0 source
• Nix current derivation entry
• Nix 2.35 derivations
• Nix build process
• Nix 2.35.2 source

Native and reproducible toolchains. LLVM and Clang define data-layout and cross/toolchain choices that must be explicit for ABI-correct code. GNU manuals document build/host/target vocabulary and semantic linker/archive/script behavior. The reproducible-builds definition supplies the correct boundary: same source, environment, instructions, and dependencies—not merely “same compiler source.”

• LLVM language-reference data layout
• Clang cross compilation
• Clang toolchain
• Autoconf target triplets
• GCC language standards/runtime implications
• GCC link options
• GNU linker and linker scripts
• Reproducible Builds definition

15.3 Repository and empirical evidence record

The source audit covered the Makefile; C and self-hosted drivers; compiler entry and both interface writers; package/test coordinator; language and test specs; bootstrap documents/recipes; and package, separate-compilation, incremental, byte-identity, and bootstrap tests. The audit found documentation drift as well as code defects: the root instructions count five unit suites while Make lists six, and the test-system document describes three pinned data byte divergences while the current Make divergence list is empty. Neither drift item influences the architecture, but both require cleanup during migration.

Raw commands, timings, traces, hashes, fixture logs, environment capture, and interpretation for section 11.4 are retained in the session checkpoint under empirical/RESULTS.md and empirical/raw/. The experiments were deliberately small and answered only material uncertainties: actual process scheduling, direct versus transitive interface consumption, private/public/link-only blast radius, nonzero corruption, path identity, Cstage/WWstage symmetry, duplicate storage, and hidden-host execution. No toy result is treated as proof that the new architecture is complete; sections 13–14 turn each inference into a gate.

15.4 Assumptions

• Breaking import syntax, package layout, compiler flags, cache/workdir format, bootstrap artifacts, and consumer builds is authorized.
• WW can add explicit C ABI/representation annotations and change its compiler export protocol without preserving .wwi compatibility.
• Official releases can distribute or name legally usable immutable tool, sysroot, runtime, and SDK closures for each supported target.
• SHA-256 is the version-1 content primitive; every format is domain-separated and versioned so a future digest transition can be explicit.
• A first release may support fewer frozen targets than impure development targets; it may not call an impure target “hermetic.”
• Byte reproducibility covers build artifacts, not identical behavior of an external kernel, firmware, network service, or hardware.

15.5 Remaining risks and bounded experiments

These experiments are implementation gates inside the selected architecture; they do not reopen its package identity, no-script, one-graph, lock/network, direct-export, content-key, native-provider, or stage-zero choices.

1. Deep .wwe closure. Prototype the real standard-library type graph and prove that one direct export file contains every transitive layout/type fact required without leaking non-semantic source data. Adjust record layout, not direct-dependency semantics.
2. Generated C seed. Measure generated ww0.c size, C99 portability, host-C variance, and fixed-point convergence on at least three unrelated C implementations. Restrict/repair the recovery emitter rather than introducing a second maintained compiler or opaque-permanent binary seed.
3. External linker/tool bundles. Certify deterministic debug info, build ID, archives, scripts, shared-loader metadata, redistribution rights, and resource closure. Change tool adapters/bundle membership if needed; do not restore ambient driver defaults.
4. Sandbox portability. Implement denial conformance on every official B platform, including process children, filesystem race/symlink attacks, clock, randomness, and network namespaces. A platform that cannot enforce it remains non-frozen rather than gaining an undeclared exception.
5. Native-provider coverage. Exercise ELF first, then Mach-O frameworks/SDKs, PE/COFF import libraries, symbol versioning, and kernel image builders. Extend the closed typed schema by version where genuinely required; do not add raw search or a general build language.
6. Registry/private-source protocol. Test mirror failover, redirects, credential isolation, key rotation, provenance, yanked releases, and malicious archives. Vendored/exact-origin operation remains the deterministic fallback.
7. Performance. Validate compiler worker strategy and CAS hashing against the measured 15-package/full-toolchain budgets. Process topology may change while action boundaries and keys remain fixed.

15.6 Final decision trace

The documented Pike evidence shapes the design: computable direct dependencies, fast compilation, cycle rejection, package boundaries, and orthogonal concepts; the collective Plan 9 papers add transparent encodings and system-wide placement of complexity. Later Go work demonstrates one possible module/cache/toolchain evolution but does not decide WW's answer. This document then applies those principles to requirements early Go's package/build model did not expose completely: foreign ABI contracts, C headers, external assembly and objects, archive/link ordering, linker scripts, CRT/libc/sysroot identity, host generators, and explicit build/host/target closures.

The resulting binding conclusion is singular: replace the current system with WW Action Build exactly as specified above. Do not preserve the old path, and do not substitute a programmable project framework or a source-only package command during implementation.