ww/docs/test-system-v2.md

WW test architecture

Status: active architecture as of 2026-08-07. The target graph, corpus accounting, and complete test-commit gate have been rechecked, including its two concurrency controls separately and together. The separately gated byte-identity and test-infrastructure proofs have also run. Bootstrap and platform proofs have not run as part of this revision.

Why the old suite was slow

The former suite made shell, Make, and one C executable per legacy case share ownership of discovery, registration, phase selection, scheduling, result records, timeout classification, and a last-green cache. Language behavior and byte-identity loops repeatedly rebuilt complete package graphs. The measured serial result was 411.403 seconds wall time, with a static lower bound of about 11,040 compiler-stage processes and 4,800 archive creations.

The replacement assigns each assertion to one owner and keeps expensive proof categories out of the ordinary developer target.

Owners

Category	Owner
Arena, lexer, parser, checker, module-decl parse, narrow codegen	Six in-process C unit binaries
Compile success/rejection, stage-routed diagnostics, and runtime exit	test/wcc/data/*/case.ww, executed by wwfixture
Package semantics	test/package, test/sep/sepinit_test.ww, and the native package-test coordinator
Language behavior	test/lang/*_test.ww through the language @test runtime
Library behavior	The package coordinator's lib/... tree walk (ww test -j N lib/...)
Standalone library-source compilation	Three import-free real source paths named by LIBRARY_STANDALONE_SOURCES, compiled directly by both frontends
Compiler-output identity	test-lang-byteid and test-data-byteid plus residual byte/artifact carriers
Fixed point and self-host	test-bootstrap
Host linker/platform behavior	test-platform

The live declarative compiler corpus has 1,759 fixtures and 3,518 C/WW cells: 351 expected rejections, 22 compile-only successes, 209 exit-zero programs, and 1,177 explicit-exit programs.

15 native C carriers remain. They are partitioned exactly once into the two terminal classes: six in-process units (738_module_decl re-homed from the residuals — Go does not distinguish "bootstrap observers" from unit tests, and its parsefile-call shape is exactly the 200_parse class), six bootstrap gates, and three platform-claim carriers (996_dyn_ww under test-platform; 810_dyn and 989_dynentry_run hold host loader/glibc claims and run with test-compiler). The residual class is EMPTY: 794_xmod_ident_prefer — the last bug-pinned survivor — retired with its #55-sibling cgen fix (its assertions migrated to test/xmod/collide_test.ww).

The 2026-08-08 residual drain retired the other 107: 5 deleted against verified owners, 196 rows migrated into corpus fixtures (700_e2e's 136, the fixture-class 46, the 926/929/671 run legs), and the rest ported to ww-native observer suites on test/testenv, each carrier retired in the same commit as its replacement: test/sep/ (sep-layout/driver), test/xmod/ (module-tree collisions, mangle, typecheck, direnum), test/asm/ (positional asm-window needle observers), test/object/ (w6a/w6l ELF, archive, and link facts), test/misc/ (divergence pins, stamp probes, wwi round-trips), test/tool/ (driver CLI, wwdump gates, FFI, the c6 soak), and test/libenv/ (the lib/os + lib/dirs env-arranger legs; those lib suites now pass bare with loud SKIPs). All suites run under test-compiler. The byte/artifact partition stays EMPTY under test/byteid/ exactly as before; wwstage-driver-leg byte identity keeps its one owner, test/byteid/libbyteid_test.ww (44-entry roster plus the lib/ completeness scan). Known cs≠ww divergences exposed by row migration are pinned in DATABYTEID_DIVERGED (3 entries) on the graduation discipline.

Carrier endgame

The declarative corpus and the native @test owners are the permanent test surface — the go/test/ analogy. The native C carrier fleet is the pre-Go-1.5 artifact and shrinks toward exactly two terminal classes:

1. C-bootstrap observers. The six in-process units watch the C frontend from inside its own process and are irreplaceable while cmd/ is the live frontend. The six bootstrap gates are the same class's proof column: they compare the wwstage tools against the C stage to a fixed point. At the eventual selfhost flip both freeze into a bootstrap smoke gate (Go 1.5 deleted its C toolchain and that toolchain's tests; it did not port them).
2. Host ABI/platform gates. Behavior owned by the host linker, loader, or ABI (996_dyn_ww today). These observe the platform, not the compiler, and stay native exactly as long as the claim is about the platform.

Everything else gets a ww owner. Byte and artifact observations — assembly-pattern greps, symbol tables, frame layouts, .wwi round-trips, driver-leg comparisons — are subprocess plus file IO plus string search, which test/package/package_test.ww already performs natively (runcommand + in-language assertions). A C carrier whose assertions fit that shape is ported and retired in the same commit, assertions preserved or strengthened, with no compatibility execution path left behind. A carrier that is merely historical is deleted outright; git history is the archive.

Public targets

Target	Composition
test	Six in-process units plus one compile-only C/WW compiler-fixture smoke case
test-compiler	Complete fixture corpus plus residual compiler/integration carriers
test-package	Package planning, grouping, routing, and package runtime only
test-lang	Language-owned @test behavior
test-library	The coordinator's lib/... walk plus three import-free standalone-source C/WW compilation checks
test-commit	Unit + compiler + package + language + library behavior
test-byteid	Compiler-output identity gates
test-bootstrap	Fixed-point bootstrap plus the 950/991–995 native gates
test-platform	Host-dependent dynamic-link gate
test-wwfixture	Fixture CLI/process/protocol integration boundary
test-all	Commit + byte-ID + bootstrap + platform + test-infrastructure checks

test-commit deliberately excludes byte identity, bootstrap, and platform work. test is intentionally smaller than the old target and is the ordinary developer feedback gate. Its purpose is a short, direct path from a compiler edit to useful evidence, not compliance with an arbitrary wall-clock cutoff.

Make and fixture scheduling have separate, explicit owners. The Makefile does not detect CPU count or add -j to MAKEFLAGS; the caller selects Make parallelism with the standard make -jN option. JOBS ?= 1 controls only the wwfixture -j N value passed by test-compiler. A normal fast gate can use both layers deliberately:

make -j4 JOBS=4 test-commit

For deterministic failure reproduction, make both layers serial explicitly:

make -j1 JOBS=1 test-commit

Serial execution is a debugging mode, not a correctness requirement. JOBS is not inferred from MAKEFLAGS, and there is no jobserver adapter or second scheduler hidden in Make.

nocc remains the separate, explicit reproduction route from a checked-in stage-0 snapshot. It is not an implicit prerequisite of ordinary tests or of test-bootstrap, because it has an external stage-0 precondition.

Compiler fixtures

Each fixture is one directory with one case.ww. Its first line is exactly one of:

//ww:error "required diagnostic fragment"
//ww:error c "C-stage fragment" ww "WW-stage fragment"
//ww:compile
//ww:run
//ww:run-exit N

• error requires normal nonzero frontend termination and the declared stderr fragment. The labeled form routes distinct fragments to the C and WW cells; labels are fixed-order and neither fragment is treated as a shared fallback.
• compile requires frontend exit zero and produces no executable.
• run builds and requires normal program exit zero.
• run-exit N builds and requires normal program exit N.

Every fixture is run against the C and WW frontends. Signals, launch failures, timeouts, build failures, and runtime exits are distinct outcomes. test-compiler passes -j $(JOBS) and therefore uses one fixture slot by default. The direct wwfixture CLI retains its own four-slot default; pass -j N when its concurrency must be explicit. Its existing os.exec.start/poll loop supervises the independent processes. Each cell has its own working directory; filesystem fixtures use that directory or an existing temp.named path rather than a shared fixed pathname.

The old TESTS list and all explicit per-wrapper Make rules are gone. A surviving C carrier is registered only by its source file and built through one generic pattern rule. Four arena/frontend units share a static rule; the codegen unit links the existing private cgen and text-emitter objects directly.

test/wwfixture/integration.sh remains the direct black-box owner for behavior that exists only at the command/process/protocol boundary: filtering and list output, phase and outcome classification, diagnostic routing, malformed-corpus and identity-drift rejection, signal/timeout/interruption cleanup, publication failure, and strict result-stream decoding. Semantic fixtures cannot prove those observations about their coordinator. test/wwfixture/process/main.ww owns the lower-level os.exec primitives, not the CLI policy layered over them, so it is complementary rather than a duplicate owner.

The six unit sources have zero active system, popen, fork, or exec calls. Before direct conversion they contained two popen call sites and a full test-unit run launched one ww -V plus twelve w6c processes. 400_w6c now checks the same twelve assembly fragments in process, grouped under ten unique sources. The CLI version assertion lives with C/WW-driver parity in the existing 949_driver_flagargs integration carrier; 000_smoke retains its arena-growth and nonempty-version-constant assertions.

Package and language behavior

ww test delegates directory package requests to the native package coordinator. The coordinator owns request-pattern expansion, package grouping, same-package and external-package action selection, one canonical directory-owned product, filtering, result aggregation, and its internal temporary workspace. Test sources are exclusively *_test.ww (Go's _test.go contract): a line-leading @test declaration in any other source is rejected loudly — by both driver stages at directory enumeration and by the coordinator at source classification. A local spelling containing ... (for example ww test lib/..., Go's ./... form) is recognized by both driver stages before path resolution and walks from the directory prefix before the first wildcard. Recursive children beginning . or _ and exact testdata subtrees are pruned; child directory symlinks are not followed. A wildcard cannot consume a non-terminal exact vendor element, while an explicitly vendor-rooted pattern remains legal. Every selected test-bearing directory becomes exactly one package run, binary, and result even when both internal and external actions exist. Selected source-bearing directories without selected test files report the usual ? line after ordinary production validation, with no support, generated main, link, binary, result, or process. The coordinator first launches one driver plan for the complete command-global package/action union. After that shared build completes, -j N schedules up to N successful selected test binaries concurrently under os.exec start/poll supervision (no threads); emission stays strictly in group order, so the byte stream is identical at every -j level, and -j 1 — the default — matches the former sequential run loop exactly. Measured on the 31-package lib/... walk: 7.0s sequential, 2.4s at -j 4.

Directory test binaries are always linked under the coordinator's temporary product root. -c independently requests a caller-visible executable copy and suppresses execution. -o independently requests a copy and still runs the temporary binary unless -c is also present. Without -o, -c publishes <import-leaf>.test in the invocation directory; an output ending in / or naming an existing directory receives one such name per selected package and missing parents are created. A single non-directory output is legal for one package only. Multi-package non-directory output and duplicate visible binary names reject before tools; exact /dev/null discards every copy and permits duplicate names. The private test link still runs unless -c suppresses it; the raw single-file compatibility route also links into driver-owned private storage instead of using /dev/null as an artifact or .sepwork stem. A package with no selected test source validates production but publishes nothing and does not create an otherwise unneeded output directory. Successful -c output is silent apart from no-test reporting.

The package suite also pins the adjacent ww build output-directory branch. After normal loading, an existing directory (including a symlink to one) or an explicit spelling ending in / receives each command under its requested import-leaf name (or local directory-leaf fallback) even for exactly one selected root; a raw .ww command uses its source basename. A missing trailing-slash hierarchy is created from 0777, filtered by umask, by the request transaction. A directory or raw no-main-only selection and an import failure run no tools and do not create output. Independently selected non-main siblings in a mixed request load but have no action; command dependencies still build. Package/import rejection precedes no-main, derived destination length, duplicate command basename, and implicit default-directory preflight. If the synthesized default command basename already exists as a directory, load/graph validation precedes a zero-tool rejection and preserves that directory; a non-main package has no corresponding default output. Persistent work reuse, invalidation, failures, interruption, and concurrent drivers retain the same action and rollback rules; output form does not enter package/import/action identity. This build branch starts no runtime process, while direct execution of its published binary remains an artifact check rather than part of ww build. Assembly-only -S still selects command actions and rejects a no-command directory request, but it performs no install-only destination validation or output-directory creation.

The retained executable is byte-identical to the temporary runnable and has executable mode 0777 filtered by the caller's umask, but it is never the path executed by the coordinator. Compile-only publication participates in the driver's one request-wide transaction: producer, linker, staging, or installation failure preserves every prior destination and removes stages and newly created output prefixes. A running retained request instead withholds the public path from that build transaction. The coordinator runs the private binary, then invokes the selected stage driver's guarded install action only after a successful process result. Failure, signal, timeout, interruption, or child-start failure therefore preserves any prior retained binary and creates no new one. Successful products install independently after their runs. The language runtime owns individual @test functions.

Every build or retained-test public install follows Go 1.26.5's late destination safety rule. After applicable producers (and, for running tests, after the successful run), ordinary stat rejects a directory and rejects a nonempty regular non-object file. Empty reservations, recognized prior outputs, and non-directory non-regular paths remain replaceable. Recognition uses Go's archive/ELF/Mach-O/PE/Plan 9/WASM/XCOFF magic plus WW's narrow //ww:module interface prefix. Package/import rejection keeps its earlier diagnostic precedence; exact /dev/null and -S never enter the guard.

Every actually executed directory product gives its single generated binary the product's canonical absolute physical source directory as child cwd. A fresh per-run environment removes inherited exact uppercase PWD entries and appends one PWD with that same directory, matching Go 1.26.5's observable AppendPWD plus last-value-wins os/exec behavior. The coordinator's cwd and environment do not change, and overlapping -j N children share no writable environment vector. Relative ordinary data, testdata, and writes resolve in the tested directory. Production and test-only dependency initializers run in that product process and see its directory; separately testing the dependency creates a separate process using the dependency directory. Equivalent direct, recursive, redundant, absolute, and root-symlink spellings converge before this runtime field is assigned.

Every such product also receives an independently opened null device as fd 0. Caller terminal, pipe, and file bytes remain with the coordinator; serial and parallel products observe immediate EOF rather than consuming a shared input offset. Production and test-only dependency initialization, filters, list mode, no-match execution, failure, and timeout use that same process boundary.

The same product maps fd 1 and fd 2 to one open product-local capture. This matches Go's test-command use of one writer for exec.Cmd.Stdout and exec.Cmd.Stderr: writes are not drained into two files and regrouped later. The coordinator emits the completed combined bytes and any product run-status line on stdout. Each parallel product has a distinct capture, while canonical group-order emission remains byte-stable across -j values. Build-plan stdout and stderr stay distinct, and loader/build diagnostics stay on stderr.

The physical directory remains distinct from exact dotted package identity and from production, internal, external, recompiled, support, and generated-main action identity. Only product execution uses it. Compiler, assembler, archiver, linker, support, and generated-main commands retain their build-plan cwd and environment. Captured directory build plans begin with null stdin, which their inherited-stdio tool descendants retain. ww build, ww test -c, -c -o, and no-selected-test products start no test child. A published test binary run directly and the raw single-file compatibility path inherit the user's cwd/environment/stdin and contain no forced package-directory or input behavior.

Before package grouping, both directory drivers and the shared recursive coordinator apply Go 1.26.5's filename OS/architecture rule for WW's fixed linux/amd64 target. The basename stem ends at its first dot; a final _test token is removed for matching; a recognized OS/architecture pair takes precedence over a recognized final single token; and unknown or misplaced tokens remain ordinary. Wrong-target production, same-package test, and external-test files therefore create no imports, variants, generated-main inputs, runtime tests, artifacts, or persistent invalidation. Names are byte-sorted before selected-source validation, so diagnostics do not depend on directory entry order. A recursive pattern omits a directory with no eligible source, while an explicit wrong-target-only build rejects it as having no WW package source. There is no source-level build-expression or user-tag mode.

After that eligibility boundary and the coordinator's required package-clause classification and production @test validation parses, the delegated loader performs selected-basename Go 1.26.5 simple-fold preflight before its graph import scan. During graph loading, effective canonical imports register only after contextual local/vendor expansion; exact identity reuse is accepted, but two distinct fold-equivalent identities reject the entire shared plan. Selected basenames register in one canonical-directory scope. A production build registers production names only. A test plan shares the filename scope across production, internal-test, and external-test actions, including same-only, external-only, and mixed directories, without combining their source units. Hidden, underscore-prefixed, wrong-platform, and otherwise ineligible files never register. Package variants, product-scoped recompiled copies, support, and generated main remain action wiring over exact identities and do not manufacture package collisions. The fold preflight does not replace or precede those coordinator-owned source-validation parses.

Separate compilation is the only driver build path; no compatibility mode switch remains.

test/package/package_test.ww owns the focused dual-stage single_root_build_output_directory observer. It covers existing and missing directory spellings; trailing and non-trailing forms; literal, logical, default-dot, symlink, raw-file, and implicit-default routes; build/test adjacency; mixed and repeated roots; command-only publication, no-main and implicit-collision rejection, non-main default behavior; import diagnostic precedence over no-main, derived-path, duplicate-destination, and recursive implicit-collision checks; -S -w command selection with long, duplicate, implicit, and missing-directory publication preflight omitted; cold, warm, and invalidated persistence; compiler, assembler, and linker action traces; compiler/linker failure and linker signal; prior-state and directory-creation rollback; concurrent isolation; executable modes and runtime exit values; .new/transaction cleanup; and Cstage/WWstage diagnostic, binary, assembly, and persistent semantic-artifact byte parity. General output permissions, occupied stages, multi-product transactions, null discard, and test runtime timeout remain with their existing observers.

test/package/package_test.ww also owns the dual-stage declared_name_identity_and_file_import_scope and explicit_import_alias_binding_modes observers. They generate temporary directory trees proving that canonical import identity, physical directory, declared package name, optional source alias, and effective source-file qualifier remain distinct. The alias matrix covers default and explicit-only qualification, bare function/type/def/const/variable rejection with exact unused-before-undefined diagnostics, same-file duplicate bindings, accepted same-path imports under distinct bindings, alias reuse across files, sibling scope isolation, package-declaration collisions, unused aliases, blank side-effect imports, repeated blank/default/explicit combinations, and blank no-binding/no-unused behavior. It also pins canonical .wwi spelling, archive/action ownership, one canonical direct edge for repeated occurrences, stage-equal diagnostics and binaries, and clean rejection state. The declared-name observer proves imports are file-scoped while dependency edges are the package-wide sorted union; command and production/internal/external/directory-main actions retain canonical action ownership while one canonical directory owns their product; external self-imports bind the augmented internal action; vendor expansion changes identity but not the effective qualifier; compiler argv contains only direct .wwi inputs; and its named rejected actions leave neither committed nor staged action artifacts or a published binary. It also forces a request transaction to reject after dependency work has staged, then proves that both stages restore every prior artifact, unit, tool record, stamp, and product rather than accepting a mixed warm generation. Together with the existing directory, recursive, vendor, exact-argv, command, and persistent-workdir observers, the package suite proves archive-only link argv and exact warm/rejection-state behavior without duplicating those broader mechanisms in this observer.

The same package owner contains the focused directory_test_execution_working_directory observer. It constructs independent temporary directories from an unrelated caller cwd and compares Cstage/WWstage output for production/internal/external/combined and every test-only shape; production and test-only dependency initialization; recompiled external self-import; duplicate inherited PWD; ordinary data, testdata, relative writes, and deliberately nonempty caller stdin; direct/recursive/redundant/absolute/symlink roots; reversed request and creation order; serial and parallel products; filters, list, and no-match execution; failure and timeout; build/no-test and compile-only paths; direct published binaries and raw single-file execution; build-tool cwd/argv/environment/stdin; persistent data-only reuse and artifact bytes; input-open failure before capture creation; source-class rejection with empty workdirs; and deterministic post-build child-chdir failure isolated from a successful sibling. It pads the inherited environment beyond former fixed observer sizes, requires one appended product PWD, requires EOF for captured actions and caller data for inherited- stdio routes, alternates fd-1/fd-2 write syscalls through every executing variant and initializer, requires one ordered stdout stream on success, assertion failure, signal, timeout, and setup failure, requires direct/raw inherited routes to keep their streams separate, and sweeps the persistent workdir for staged residue. Command-global bounded-memory failure remains independently owned by allocation_failure_is_command_global.

The same package owner contains the focused platform_filename_source_selection observer. It independently generates the suffix matrix, production/internal/external sources, wrong-target import and malformed sentinels, wrong-target-only directories, and reverse-created diagnostic files. Across cold and persistent Cstage/WWstage roots it compares normalized compiler/assembler/linker argv, units, .wwi, assembly, objects, archives, generated-main archives, binaries, runtime/test output, direct and recursive behavior, reversed roots, ignored-edit reuse, selected private-change propagation, and request rollback after a dependency has staged and the root compiler fails. Wrong-target test files cannot add a target to the directory main or create one for an otherwise no-test directory. The bounded-memory package-initialization observer now pressures the combined directory topology itself: production, augmented internal test, external self-import, a transitive recompile-for-test clone, support, and the two-target generated main. Each stage searches its own bounded address-space ceiling, while the accepted failure must have the same ww: out of memory diagnostic, invoke no compiler, assembler, or linker, and leave an empty caller work directory with no output, status, or staging path.

The same observer now owns selected-basename folding: ASCII production, internal, external, same-only, external-only, mixed, and cross-classification collisions; printable Unicode Kelvin folding; malformed UTF-8 replacement-rune keys with exact \xNN diagnostics; accepted composed/decomposed neighbors; ordinary-build isolation from test collisions; ignored folded pairs; reversed creation order; and direct/recursive collision-diagnostic parity. It proves zero compiler/assembler/linker calls and an empty workdir, including no test-support/main artifact, on cold rejection. A committed warm package then gains and loses a colliding file; the failed middle request preserves every unit, .wwi, assembly, object, archive, binary, tool record, and stamp, and removal restores exact reuse. package_graph_diagnostics_are_stable provides the corresponding direct, transitive, reversed-root, recursive, same-directory, exact-repeat, and warm import proof. vendor_directory_import_resolution proves folding uses the expanded vendor identity under reversed imports and product descriptors, even when both spellings reach one physical directory, and that multi-product rejection publishes no output, status, tool state, or staging residue.

test/sep/sepinit_test.ww is the single focused package-initialization owner. It generates all source trees temporarily and runs independent cold/persistent Cstage and WWstage legs. Its matrix covers blank-only reachability; dependency, diamond, and independent ready-task order; runtime call/allocation and aggregate package lets; multiple init declarations and invalid forms; direct/qualified init invisibility; initialization cycles and diagnostic order; command, internal-test, external-test, support, and one directory-generated-main ownership; canonical task and dispatcher symbols; deterministic one- and two-member archives; .wwi exclusion; byte-identical artifacts/binaries; init-only invalidation; warm invalid-init rollback; direct/recursive variant equivalence; exact production, support, ptest-before-pxtest-before-main task ordering, exactly-once package/dependency initialization in one dispatcher; complete normalized rejection diagnostics; dangling staging/rollback no-follow rejection; and checked compiler-output failure/non-regular-destination rollback; and bounded-memory Cstage/WWstage allocation-failure parity across the complete combined package-test graph, under independently discovered ceilings supplied by the repository-built sep-limitexec helper.

ww build and an explicit single-file ww test -o <stem> publish <stem>.sepwork as a caller-owned artifact directory. The driver acquires it with one fresh mkdir and refuses an existing path; it never clears a collision. A caller keeps only the exact artifacts it observes and removes that exact tree on every later success or failure. Directory-package test plans instead keep their cold semantic-action scratch and runnable binary inside the coordinator's temporary root; a compile-only retained executable escapes through the build transaction, while a running retained executable uses the post-run install action above. ww run and no-output single-file ww test use driver-owned scratch instead; both driver stages place that scratch and their temporary executable beneath one freshly acquired directory, remove both after every build result, and make cleanup failure fail the command. Make recipes build driver-produced tools in invocation-owned directories and apply the same exact cleanup rule.

ww build -w DIR and single-file ww test -w DIR replace that scratch with a caller-owned persistent package-artifact workdir: for these direct routes the directory must already exist, is never cleaned by the driver, and holds one committed unit, .wwi, .s, .o, and dep .a per package plus byte copies of the invoking driver, compiler, and assembler and a small mode stamp. A package is reused only when its freshly composed owner unit byte-equals the committed unit, no recompiled direct dependency emitted changed export bytes, and every applicable recorded executable byte-equals the live executable — content identity only, no mtimes, no hashes, every decision reproducible with cmp against plain files. The driver identity covers graph construction, owner-unit composition, direct-export argument construction, archive creation, and commit algorithms that neither unit bytes nor compiler identity can name. Recompiled artifacts, dispatcher artifacts, products, statuses, tool copies, and the stamp land at staged .new names. After every requested product stages successfully, one rollback-capable transaction installs the whole generation; any producer or installation failure preserves the prior committed bytes and removes remaining stages. An interrupted or rejected request therefore cannot create false reuse or a mixed generation; a successful executable link still reruns when required. One workdir serves one invocation at a time and may hold the complete command-global action universe for all selected roots and isolated variants; semantic identity, rather than request shape, controls reuse. Both driver stages implement the identical contract. Test-mode persistent workdirs use format 19. This is build staleness in the Make/mk/Go sense, not a result cache: real directory test products always run, while no-selected-test directories do not create a process. The byte-identity and bootstrap gates keep building on fresh scratch. make clean reclaims every workdir under out/.

Directory-product cwd and PWD are request-time child-process metadata, not persisted action inputs or results. A runtime directory-entry failure therefore does not erase a successfully committed compilation generation. Changing only ordinary fixture data causes no unit/export/assembly/object/archive/main/binary change and no producer work beyond the same established warm final relink, while the next always-run product observes the new bytes. Build workdir format remains 18, test workdir format remains 19, and semantic storage remains 3.

Fold keys and first-spelling tables are request-only loader state and are never stored in a workdir. They do not alter successful unit bytes, action/storage keys, or tool records, so build format 18, test format 19, and semantic storage format 3 remain unchanged. Structural fold rejection occurs before staging and cannot mutate a committed generation; a later request with the collision removed evaluates the original exact-key reuse contract.

On a package or tree target, ww build/test -w DIR forwards that exact caller path as the one command-global semantic-action store; neither pattern spelling nor package-group name derives a persistent subdirectory. The delegated driver may create a missing store only after graph and request preflight, and rolls back newly created empty prefixes if later setup fails. Every reuse decision stays with the driver's content-identity contract above. Test -c may use the same store: the workdir owns semantic actions and the invocation/output path owns the separately staged retained executable. The test-library route uses the same command-global store contract: one ww test -j $(JOBS) -w out/wwbuild/wwtest-lib lib/... line owns library behavior (measured: 2.3s cold, 1.0s warm at -j 4, vs 4.7s for the retired 59-target per-file fan-out).

test/lang currently uses one package per source file, so its complete gate still performs independent package builds. That remaining source layout is not hidden behind caching or concurrency; it is outside the small test target.

Compiler-only byte identity

ww build -S and ww test -S -o <stem> run source discovery, dependency ordering, unit composition, and each required w6c -c invocation. They return after the complete .s/.wwi set exists. The producer loop does not invoke w6a, create per-package archives, or invoke w6l when -S is active.

test-lang-byteid runs the 158 selected language files once with C w6c and once with WW w6c, requires identical emitted .s filename sets including __root.s, and compares those bytes. Relative to the old two-leg full builds, this removes at least 3,476 assembler launches, 3,160 archive writes, and 316 linker launches from the comparison loop. A cold Make invocation may still build prerequisite compiler binaries; the compiler-only claim applies to the per-language-file comparison path.

test-data-byteid applies the same comparator to the declarative corpus: every non-error case.ww builds twice through the fixed cstage driver with only WW_W6C swapped, and every emitted per-package .s must be byte-identical. //ww:error fixtures have no .s; their both-stage reject parity is owned by the fixture corpus itself. Known cs/ww divergences are pinned in DATABYTEID_DIVERGED with the 989_lib_byteid discipline: a pinned fixture must still build on both stages and still differ, so a compiler fix fails the gate demanding graduation rather than silently widening coverage. The full sweep compares the 1,157 non-error fixtures in about a minute and is scratch-rooted under out/, not /tmp.

Byte identity is an explicit proof gate. It is not a prerequisite of test or test-commit.

Bootstrap, subprocesses, and CSP

Stage-2-through-stage-4 fixed-point proofs and the 950/991–995 self-host gates are reachable through test-bootstrap and test-all, never through test or test-commit. Cold ordinary targets may still build their C- and WW-stage tool prerequisites once; they do not iterate those tools to a fixed point. The C bootstrap source and the standalone nocc route remain intact.

The bootstrap recipe owns the fixed out/bootstrap tree. Make schedules that target once within one invocation, but two independent make bootstrap invocations are not safe to run concurrently and remain mutually exclusive.

lib/os/exec is the sole reusable WW subprocess mechanism. Fixture and package coordinators use its captured asynchronous path. An empty captured-command stdinpath opens the null device; an explicit path supplies controlled input, and either descriptor is installed before exec with checked setup reporting. Distinct output paths create independent exclusive captures; byte-equal output paths open once and duplicate that descriptor so fd 1 and fd 2 share the same open file description. The executor process proof covers both modes, including closed caller standard descriptors. The WW driver directly uses os.exec.runstdio for inherited-stdio, inherited-environment, leader-only compiler, assembler, linker, cleanup, run, and single-file-test calls. The local WW procrun implementation is deleted. The C bootstrap retains its C process implementation because it cannot consume a WW standard-library module.

WW has tokens and an opaque type for future CSP/channel work, but no mature production channel operations, task runtime, or scheduler. No channel, goroutine, thread, worker-runtime, or CSP library was added. Coordinators stay single-threaded; OS process polling does not require language-level threads.

Retired mechanisms

The following are deleted, not adapted:

• test/run, including its phase classifier, xargs scheduler, atomic private records, timeout-text classifier, result collector, and last-green cache;
• test/run_test.sh, the synthetic shell tests for that protocol;
• every explicit TESTS registration and all 329 explicit wrapper rules;
• the unregistered test/runww.ww corpus runner;
• frozen duplicate compiler-corpus code under internal/wwtest, test/wwtest, and test/compiler;
• the redundant standalone smoke, test-run, and test-harness routes; and
• library-launcher wrappers whose only assertion was an existing @test source's exit status, including the declarative 900_stdlib.c launcher.

There is no test cache, daemon, scanner, generated manifest, database, new framework, compatibility API, concurrency runtime, dependency, or changed timeout policy in this architecture.

Open driver work

None; directory-package -c and -o now have the applicable Go 1.26.5 retention, naming, fan-out, execution, and publication behavior.

Validation policy

Use test-unit as the inner loop for lexer, parser, checker, and narrow codegen changes. Run test for the ordinary local compiler check, followed by the focused owner for the changed behavior. Use test-commit for ordinary pre-commit behavior; use make -j4 JOBS=4 test-commit when parallel feedback is desired, and make -j1 JOBS=1 ... to reproduce failures deterministically. Run test-byteid and test-bootstrap only when those proof categories are intended. test-all is the exhaustive CI/release composition and should not be launched casually.