cgindex's tagged-element classification whitelisted base node kinds;
call- and slice-based tagged elements fell off the list and dropped
the payload words (review finding #23). Classify by the stamped
element type. 989_taggedidx_run pins cs==ww (red 2/6 pre-fix).
A bare module-global unsigned operand got signed IDIV/SAR/Jcc — the
predicate's ident arm only consulted the local table, so globals fell
through to signed (review finding #25). Read the stamp for the global
arm; corpus emission is unmoved (no bootstrap code div/shift/cmps a
bare unsigned global). 989_gunsigned_run pins cs==ww (red 3/8
pre-fix).
for-range destructure of an element with a slice- or tuple-typed field
strode by the default 8 (paramfieldsize had no N_TSLICE/N_TTUPLE arms),
silently reading the wrong words (review finding #43; live repro
cs=42 vs ww=8). Add the arms routed through tinfo per rule 13.
989_tupfieldsize_run pins cs==ww (red 1/2 pre-fix); rows assert
convergence, not absolutes — cstage's own single-word destructure-load
bug is filed as task #40. The N_TARRAY arm is deferred (task #39,
rule-7 comment at the fall-through).
cgindex read one word for a str/slice ELEMENT of a chained index
(xs[i][j], f().s[i]) — the element-kind gate keyed on node shape and
missed non-simple bases, dropping the 24B/16B header load (review
finding #22). Key on the element-type stamp; 989_chainidx_run pins
cs==ww (red 3/8 pre-fix).
A slice/str ELEMENT of an indexed expression passed as a call-arg
pushed one word instead of the 24B/16B header — the predicates had no
N_INDEX arm, so element-typed args fell to the scalar path (review
findings #45/#46). Read the element-type stamp; dual-stage rows in
989_idxarg_run pin cs==ww (red 2/8 on pre-fix binaries).
Collapse the localfindnode tnode read to the stamped n.type_ — a
zero-delta mechanic validation for the F7 predicate type-keying arc:
self-compile byte-id holds on all four bootstrap combineds, proving
n.type_ matches the local-table truth on every corpus local before
any behavior-changing conversion lands. The localfindnode gate stays
so a global-not-def ident still routes false.
wwstage's desugarcallargs ran no general per-arg typecheck (only the
narrow #258 array-to-slice arm): any mistyped scalar call-arg silently
miscompiled (int read as a 24B slice header; the -T face was a user
const __wwtests building a garbage test binary). Route every call-arg
through the predicate union isassignable()||assignableaddrfn(),
mirroring cstage type_assignable||assignable_addrfn and the check.c:1869
diagnostic. Confident scalar/aggregate and aggregate/aggregate
kind-mismatch rejects live in shared isassignable; the concrete-to-
tagged arm is shape-matched-lenient via tagshape() (AST mirror of cgen
taggedvariantindext) so genuine variant members keep flowing while
shape-mismatched aggregates reject. Reserve __wwtests under -T in both
stages (mirror the main reservation, check.c:2996). New table-driven
989_callarg_typecheck, 31 fixtures, reject rows proven red on pre-fix
binaries.
Deferred, filed, site-commented: the assign seam rides #178->#36
(typeeqast cannot compare variadic/module-qualified fn sigs); the
same-coarse-shape same-leaf nominal collision over-accept rides #37
(#10/#66 — the distinguishing module is absent from the AST surface
isassignable operates on).
wwstage stamped N_RUNELIT as concrete rune (cstage: untyped_rune,
check.c:1296), so an uncast rune literal into a fitting integer slot
over-rejected where harec accepts (promote_flexible). Add coercerunelit
(mirror of coercefloatlit) at the scalar seams — let-init, return,
call-arg, index-store; coarse accept, no range check (mirror cstage
type.c:379); array-literal elements keep the range gate
(arrlit_init_fits, types.c:923). Teeth: the fnmatch.ww '\\': u8
workaround cast and getopttest's 'X': u8 index-store casts are removed
and compile green. Checker-only (no e.type_ restamp): byte-id held.
For an SK_FN sym, exprtype's N_IDENT and N_DOT arms returned decl.lhs
(the return type), so a bare fn rvalue compared as its return type —
latent silent mis-accept of mismatched-signature fn assignment (the
706/768 over-reject under any general check was the visible face).
Synthesize the N_TFN instead (mirror of assignableaddrfn's synth), and
add the N_TFN-vs-N_TFN confident-mismatch reject in isassignable.
harec: resolve_function check.c:4279-4288 (obj->type=fntype),
EXP_ACCESS check.c:341-343 (bare ident yields obj->type, no decay),
call check.c:1573, assignability types.c:1001; cstage twin
check.c:1313 + build_fn_type check.c:2611. cgen is name-keyed on fn
rvalues (never reads the stamp): zero asm delta, byte-id held.
White-box @tests move into a colocated 'package regex' file (Hare
+test.ha analogue; #6 non-T drop plays the build-tag role), black-box
@tests stay in package regex_test — Go's foo/foo_test split. The
wb/ stub driver is documented-temporary scaffolding: a same-dir
importer file-resolves to regex.ww before dir-enumeration, so only a
different-dir import bundles the whitebox sibling (task #33 retires
it). CLAUDE.md rule-9 carve-out updated (#5 closed, gate was #6).
`ww test <file> <pattern>` runs only the @test fns whose names match the
fnmatch glob; no pattern runs all (byte-for-byte the pre-filter path);
zero matches prints "No tests run" and exits 0 (Hare ground truth
ref/hare/test/+test.ha:114-117). A pattern in directory mode is rejected
"ww test: pattern needs a single test file" (rc 2), identical wording in
both twins (cmd/ww/main.c do_test + selfhost/cmd/ww/main.ww dotest).
Mechanism (a): rt/start.s stashes argc/argv into rt_argc/rt_argv getters
(rt_envp twin shape, -T synth untouched so 990-997 byte-id holds);
lib/os.args() rebuilds the []str view, build-once-cached; lib/test/run.ww
imports fnmatch and filters av[1..] (argv[0] is the binary path). The
driver forwards the 2nd positional as argv[1] via fork/execv (cstage) /
procrun (wwstage) so glob metachars aren't shell-expanded.
os.args() is the first `alloc`-caller in the base os module, so os.ww now
imports rt — the `alloc` builtin's malloc lowers to rt_malloc only when
the rt binding is bundled (mirror lib/strings/strings.ww:30); without it a
plain `ww build` of any os-importing program links bare libc `malloc`
(undefined). os is bundled by ~every program, so this is load-bearing.
The lib/test floor rises os-only -> os+fnmatch+ascii+strings in every -T
build; the bundled `ascii` module vs a `@test fn ascii` collision that
exposed is closed by the preceding #30 promote commit. 989_test_filter
pins the full matrix on both twins byte-identically; 949 gains the
dir-mode reject row. (#17)
A primary-package top-level decl whose leaf also names a bundled module
(`type sym` vs `import sym`; the lib/test->fnmatch->ascii -T floor's
`ascii` module vs a `@test fn ascii`) collided in the flat scope: #23's
installtop dup-check false-fired "duplicate fn/type" where cstage
coexists. cstage keys the module name out of the value namespace by
PROMOTING the same-leaf SK_USE in place to the value kind with
use_alias=1 (cmd/wcc/check.c:2831/2848/2871/2907/2928/2951), so one
correctly-kinded sym serves bare refs (call/structlit/var) while
`name.member` still resolves the module via the `kind==SK_USE ||
use_alias` N_DOT guards (check.c:87/1337).
This REVERSES wwstage's documented two-sym coexistence design
(lib/ww/sym.ww scopelookupuselocal): keeping the SK_USE as a separate
coexisting sym ripples into every bare-reference resolver (~25
scopelookup sites), and a missed site is a byte-id-consistent-but-wrong
cat-A risk the gate cannot prove away — the same failure mode retired
with the name-keyed variant-match cluster. The promote model is correct
by construction: one sym of the right kind, identical to cstage.
sym gains a use_alias field; the three N_DOT/N_CALL module-qualified
guards honor use_alias. cstage installs every SK_USE in a dedicated
first pass, so its value-arm promote is order-INDEPENDENT; wwstage
installs in source order, so BOTH directions of the collision are
promoted to reach cstage's identical single-sym end state:
- use-before-value (`import aa` then `fn aa`): installtop promotes the
pre-installed SK_USE to the value kind (use_alias=1).
- value-before-use (`fn aa` then `import aa`): installdecl's N_USE arm
promotes the pre-installed value sym in place (set use_alias=1,
no coexisting SK_USE), mirroring cstage's self-import N_USE arm
(check.c:2823-2834 `if (prev) prev->use_alias = 1`).
Both orders compile + are cs/ww byte-identical AND byte-identical to
each other. Cannot split: promote without the guards leaves
`name.member` red on the promoted sym; the guards without promote are
inert (no use_alias is ever set) — no bisect-clean intermediate. (#30)
Pins (910/997): modfn_coexist_ok (use-before-value) AND
modfn_coexist_vbu_ok (value-before-use) both accept on cstage + are
cs/ww byte-id + run to exit 6 (bare fn and qualified module both
resolve); the dup_fn row still rejects both stages (regression). fnmatch
byte-id holds. The order-dependence is exactly what regresses silently,
so both orders are pinned.
wwstage over-rejects harec's flexible-rune-const narrow (rune lit
promotes into any int whose range fits — harec check.c:1910,
types.c promote_flexible); explicit cast until that promotion lands
in ww (#29, revert to bare '\\' then). With the enum-as fix fnmatch
is byte-identical across stages: 989 #59.14 graduates
M_WWREJECT -> M_ID, transferring 972's cstage 8/8 behavior to the
provably-identical wwstage binary — the dark-regression gate. (#29)
The value-passthrough arm gated on node shape (isenumexpr); a
constant-folded enum member reaches cgen as an int-literal node,
missed the arm, and fell into the tagged-union assertion path
(CMPQ $0 + JNE -> exit 1) — silent wrong code for every `enum as
int`, fnmatch's flag tests included. Gate on the stamped operand
type (tichase == TY_ENUM) at the cgtypeassert choke point, mirroring
cstage cgen.c N_TYPEASSERT; dead node-shape helpers deleted. Tagged
`as` assertion path byte-id-unmoved (control rows). Six @test pins
in attest_pass.ww, dual-stage. (#27b-team, cat-A)
installdecl routes all four kinds (fn/type/def/let) through installtop,
which turns scopedefineinmodule's nil return into cstage's exact
"duplicate <kind> <name>" reject, keyed (name,mod) so cross-package
same-leaf decls coexist. Builtin redecls are dropped, not dup-errored:
cstage never scopes builtins (lookup_builtin first, check.c:69), so a
user redecl is dead there — wwstage mirrors via scopesamekeysym +
no-source-decl test. -T synth __wwtests installs direct, mirroring
check.c:3079. Closes the silent dup-fn hole (user fn run vs lib/test
run built a broken test binary with no diagnostic). Per-kind reject
rows + cross-package/builtin accept byte-id rows + -T collision parity
row in 910/997. (#23-team, category-A addendum closed)
lib/test/run.ww: fork+wait4 runner; each @test runs in its own child,
abort/SEGV/FPE decoded from wait-status, failures recorded and the run
continues; exit = fail count. Tests are hermetic: module globals do not
persist test-to-test (fresh fork image; sanctioned divergence from
harec's shared-process __test_main, no setjmp/signal layer needed).
-T synth (both stages) emits a module-global (str,*fn() void) table +
return run(table) instead of straight-line calls. Driver twins bundle
lib/test under test mode and gain ww test -c/-o (go test -c) so the
byte-id gates diff the same artifact the real path builds. Gates
989/910/997 rewired onto it; new 911 pins record-and-continue across
all three fault classes; 949 +3 rows. (#17-team commit-2)
Splice @test N_FNDECLs out of the unit after the body-check passes,
mirroring harec's checked-but-not-emitted: a broken @test body still
errors loudly in non-T; @test-free units are emission-unchanged.
910/997 table rows pin keep/test x non-T/-T, head+consecutive unlink,
undef-body reject, and plain-calls-dropped loud link-fail. w6c+wwdump
combined.ww regen. (#6-team)
Lone -I/-L/-l/-o now error "ww <cmd>: -X needs an argument" (rc2),
matching selfhost main.ww. do_test rewritten to an -I-only loop that
loud-rejects -l/-L/-o/unknown instead of silently swallowing them.
949_driver_flagargs: table-driven, 10 rows, each run against both
twins asserting identical rc+stderr. (#15-team)
'ww test' gains the istest build path (-T injection in build_one/
buildone) and do_test/dotest accept -I, mirroring do_run - both twins.
The 35 converted lib tests drop their interim bare mains (-T
synthesizes the entry from @test fns and rejects a user main); their
35 C run-drivers flip 'ww run' -> 'ww test'; 989_lib_byteid compiles
lib tests under -T (8 user-main probe fixtures stay non-T, gated on
the fixture field). Abort-on-first-failure stands until the deferred
record-and-continue harness lands with the multi-package arc.
A bare 'let x: T;' with 1 <= size(T) <= 7 matched no zero-fill arm in
either stage (8B and >8B were already zeroed) - 'let c: [3]u8;' read
stack garbage. User-ruled zero-value semantics: cstage gate sz>8 ->
sz>0; wwstage zsz==8 arm hoisted above the fill-run arm (required -
8B would otherwise route into the run and diverge) and run gate
zsz>0. New 840 pin: dirty-frame probe rows, dual-dim (run + cs/ww
byte-id); discriminators fail exit-154 on pre-fix binaries.
Fused with the lib/bytes test conversion (rule 11): either half alone
turns 967 red. The old exit(signalled+10) wrapped a real 1782-count
ltrim failure to exit 0 - green depended on the garbage. Converted to
assert form (completes the 35/35 @test conversion); ltrim rows keep
the bare 'let c: [3]u8;' as the consumer proof of the fix.
bytes EXCLUDED: its green masked a real signalled corruption (filed;
conversion rides the fix). stringstest documents the per-row signalled
pinpoint loss.
signalled/fail() counters and run-loop plumbing removed; failure sites
become assert(!(cond)); table rows and @test fns 1:1. Bare mains stay
until the -T flip commit. Spec: .ai/drew-t2-conversion-spec.md.
With intermediates following the output stem, concurrent wwstage-driver
builds no longer share paths - 950/990-997 join the xargs -P group.
Measured: 202.7s -> 153.8s full-test wall (x2 runs race-clean); the
serial tail (57.7% of wall) is gone, remainder is CPU-bound. CLAUDE.md
rule 14 documents the new reality: NNN range no longer a correctness
constraint, still keys phase membership / skip set / cache set.
build_one/buildone gain an explicit objstem: 'ww build -o X' derives
main.{combined.ww,s,o} beside X; 'ww run' uses its per-pid /tmp stem;
default no--o stays next-to-source (load-bearing for the make regen of
tracked combined.ww and the freshness gate). Realizes the redirect TODO
in 995_self_rebuild.c. Kills the concurrent-test torn-read race on
selfhost/cmd/<tool>/main.* (the 991 transient). 993/995/949 pass -o
into their per-pid workdirs; 949's '-o /dev/null' relied on the old
driver ignoring -o (audited: the only live driver case). Atomic
combined write (os.rename) deferred to its own commit - lib/os lacks
rename and that addition is an import-floor regen.
The 990-997/950/989_lib_byteid gates are pure functions of (stage
binaries, tracked sources, fixtures, harness, Makefile) and eat 87% of
full-test wall (measured, .ai/testperf-baseline.md). test-commit hashes
exactly those inputs; a match with the last green full run reports the
gates as 'cached' (counted separately, never silently) and runs
everything else - 25s vs 212s on a hit, full run on any miss. A red
run never records the key; make test itself is never cached and stays
the pre-push bar (rule 14 updated). Key gap caught in review:
examples/ feeds 996_dyn_ww - included.
Emits a timing block (durations, phase walls, total) before the summary
line; the ok/FAIL and 'all N passed' contract is unchanged and the only
consumer (make's exit-code gate) is unaffected. Measured baseline at
4ff79bf: 211.6s wall, 990-995 self-compile gates = 87% of wall — the
content-keyed skip tier (T1) targets exactly that block. Baseline data:
.ai/testperf-baseline.md.
Both driver twins: an import that neither locates as a file nor is
satisfied by an inline 'package <name>' declaration in the unit is now
fatal "ww: cannot find package <name>" (was a silent skip that masked
dead imports and typos). The inline scan is a new every-line helper on
the uncapped comment-skip core - peekpackage stops at the first decl,
and single-file multi-package fixtures declare several. 949 pins both
branches (miss->fatal, inline->build+run); 993 adds ww_ww parity.
mem.ww was retired in Phase-0; these two imports survived only because
the driver silently skipped unresolvable imports. Provably dead: no mem
module exists and both units build identically without them. Removal
precedes the loud driver, which would otherwise fatal on them.
check-(c): a package importing itself (any spelling) is a hard error,
mirroring Go. Predicate is leaf==owner at the N_USE/installdecl seam —
sound only after the PREP commits (dotted-test renames, package-less
boundary directive). Identical wording both stages; diagnostics-only,
byte-id-neutral. 948 pins the reject in both compilers; 708's
pos_selfimp (which pinned the abolished self-import skip) converts to
neg_selfimp + new pos_crossmod preserving the param-shadow tolerance
the case existed for. Checks (a) unused and (b)/(d) name-membership
stay deferred to the multi-package arc: imports are filename-keyed
pulls, so those need import->file provenance this compiler lacks.
Driver twins emit the line-comment directive only before package-less
files (peekpackage==nil); both lexers tokenize it (TK_MODRESET, appended
=87 so existing token values hold) and both parsers reset curmod — a
package-less file's decls get module "" instead of inheriting the last
bundled package (the sticky-curmod leak, task #11). Withdrawn
alternative: injecting 'package main' flips non-entry symbols
bare->main-prefixed (FFI-visible, broke 764). Codegen-neutral by proof:
bare symbols preserved, both stages emit byte-identical asm for a
directive-bearing combined. Transitional until strict-package rejects
package-less files outright. Includes 737 bad-deep pin for the
PREP-peek >2048 edge + 904/toktest rows for the new token.
Pre-existing rule-10 asymmetry: 4 lib files (decimal/ftos/stof/memio)
declare package past byte 2048, so the capped wwstage peek returned nil
where cstage's uncapped line-scan found the decl — corrupting the
strict-package check and any peek consumer. Mirrors cstage peek_package
semantics exactly; prerequisite for the upcoming package-less boundary
directive to inject identically in both twins.
package L + import P.L is self-import in Go terms regardless of path
spelling (the parser drops the dotted prefix); the text-level
import-string!=package-name comparison hid these 3 from the PREP-a set.
All three verified black-box (qualifier-only exported-API use) — plain
rename, dotted imports kept as legit cross-package.
@test was parsed then dropped (no consumer); `ww test` needed a hand-written
main listing each test by hand, so adding a @test and forgetting the call
silently skipped it. -T makes the checker collect @test N_FNDECLs in source
order, loud-reject a user main, and append a synthetic
`export fn main() i32 { t0(); ...; return 0; }` at the install->body-check seam;
the existing cgfn emits it (cgen untouched) -> byte-identical by construction.
Mirrors harec's checker-side is_test placement.
Plan-9-lean reduction (user-sanctioned, reinstatable post-CSP): sequential,
abort/nonzero=fail; no setjmp isolation, no fnmatch filter, no file:line.
910/997 rewired from a regex scanner to driving `w6c -T` directly (thin trusted
drivers; the @test content stays ww), with a cross-stage byte-id assert on the
-T output. attest_userman/attest_badsig pin the user-main and bad-signature
rejects.
wwstage's tagged→tagged subset-assign arm accepted all (lenient escape), silently miscompiling implicit wide→narrow: a (int|bool|str) holding a str, assigned to (int|bool), ran the int arm and read the str pointer as int. cstage rejects loud; this escape was the lone divergence.
Replace the escape with cstage's subset walk (src ⊆ dst): a src variant is covered iff typeeqast matches (structural — []u8/nested/primitives) OR both are N_TNAME with equal leaf names. The leaf bridge covers cross-module forwards where a callee's bare inline-union variant (utf8's `done`) meets the consumer's qualified `utf8.done` — raw typeeqast can't, and the escape was masking it for every forward. Mirrors casecovers (check.ww:4136). cstage untouched (align-up); spread-bearing unions keep the escape (#199b orthogonal).
Honest floor: leaf-only defers true module identity to #10 — a callee's defining module for an inline-union return is unrecoverable at check-time (the call node is gone; #211 fnretlookupmod is cgen-only). Retained divergence = cross-module same-leaf-collision over-accept (absent from bootstrap), documented at the site and filed as #10-A7 / census cat-A.
Test 989_tagged_subset_reject is table-driven with composition-discriminating rows: []u8 subset (typeeqast-only), bare↔qualified xmod forward (leaf-only), xmod named genuine-absence (reject). 352 green; w6c unchanged, w6c_ww 4b4496b8→4b316f01.
typeeqast's N_TNAME arm compared variant types by SURFACE string
(streq(aa.str, bb.str)), so a cross-module type referenced bare
(oserror) vs qualified (os.oserror) mis-compared unequal -> A6:
concrete->tagged silently mis-identified the variant. Resolve each
name to its canonical type sym before comparing: keep the streq
fast-path, else aliassym(c,aa)==aliassym(c,bb) (ww's existing resolver
maps bare [#53] and qualified [#51] to the same SK_TYPE sym). AST
analog of cstage type.c:278 TY_NAMED a==b / harec types.c:579
ident_equal -- NO interning. `c` threaded into 21 typeeqast sites.
B-full Layer 1: closes A6/#14; Layer 2 (A7 tagged->tagged subset
reject) stays deferred to the #199b flatten arc. Byte-id-neutral;
graduates shlex #59.15 (989 M_WWREJECT->M_ID). test/wcc/839 pins
both-stage symmetry (discriminating teeth = 989 #59.15, per 839 doc).
A module-global tagged init via cast (`let g: u = true: u;`) silently
emitted tag=0; cstage louds. Route emittaggeddata through flatvariantidx
(the cstage cg_tag_for_variant twin) so an unsupported variant louds while
valid str/slice cast-inits still accept+run. test/wcc/838.
A same-type tagged cast return (`return x: u`, u == fnret) emitted tag=0
(a 2nd-variant value returned the wrong payload). Route it through the
widener (needswiden when N_CAST && ru1==fu1, beside S1's ru1!=fu1) which
peels the identity cast internally -> correct; same-type non-cast stays on
the passthrough. Byte-id-neutral. test/wcc/837.
isassignable's untyped-float (A4) and nil (A5) arms over-accepted a source
into a tagged dst where no variant accepts it -> silent over-accept then
tag-0 miscompile; cstage louds. Reject (cerr + c.errs+=1). Part of the
isassignable tagged-dst over-accept class (#23 fixed arm3; A6/A7 deferred
to the B-full nominal arc). test/wcc/836.
A widening sub-union cast (return/let `v: inner` into a union whose member
is the nested `inner`) skipped the nested widen arm and emitted tag=0 (a
2nd-variant value returned the wrong payload). cstage loud-stops #35;
wwstage now matches: cgwidentaggedstorebp louds on the subset cast, and
cgreturn routes genuine-widening tagged returns (rhs TY_TAGGED, ru1!=fu1)
through the widener choke-point so the same loud fires. Same-type returns
stay on the proven passthrough. Faithful inner-tag->outer-index remap
deferred. test/wcc/835 S1 row.
isassignable's untyped-int arm fell through *confident=false/return true,
silently accepting an untyped int into a union whose variant is itself a
nested (non-flattened) union; ww emitted tag=0 (wrong arm) where cstage
louds. Reject unless a direct variant is numeric-or-enum (N_TENUM accept
mirrors cstage type_isnum). Faithful flatten+rebox deferred (nominal
identity, post-CSP). test/wcc/835 (new) + Makefile.
A tuple whose element chases to TY_ARRAY/STRUCT/TUPLE (>8B) silently
miscompiled both stages: t.0[i] read segfaulted and construction dropped
the payload into the 8B slot. Reject the type at resolution (DISP-B);
faithful inline layout deferred to #60. cstage resolve_type N_TTUPLE
(declared) + N_TUPLE expr (inferred literal, was a cstage-only silent
miscompile + cs!=ww asymmetry); wwstage tinfofornode covers both.
test/wcc/832 + 941 migrated.
