Test fork instrumentation output reproducibility across processes

[brandonpayton]

Why

#907 already landed the production fix that sorts fork-instrumentation regions before assigning synthetic locals and instruction IDs. That makes the output deterministic, but the existing regression repeats the transform inside one process and does not vary Rust's randomized hash-map state.

This PR keeps the remaining test and documentation work:

• run wasm-fork-instrument twelve times in fresh processes and compare every output byte;
• use alternating i32/i64 nested regions so a changed allocation order is visible in the Wasm;
• document the cross-process byte-reproducibility contract and the current ABI 39.

Removed from the old branch

• the production sorting change, because Batch PHP support and bundle ABI updates with their runtime changes #907 already landed it;
• the fix/teardown-reclaim-readiness (Make teardown reclamation CI wait for a blocked worker #889) ancestry;
• stale ABI 18 wording.

The rewritten branch is one commit directly on current main. Its pre-rewrite head was f3c2c587c65ed7ed15507697efc2b381acd1a2d4.

ABI and packages

This adds a test, a WAT fixture, and documentation only. It does not change the fork format, Wasm exports, host/kernel ABI, ABI_VERSION, ABI snapshot, package inputs, package revisions, or built package bytes.

Validation

Passed through scripts/dev-shell.sh:

• bash scripts/ci-run-test-suite.sh fork-instrument — 187 tests passed;
• the exact fresh-process regression by itself — 1 test passed, exercising 12 separate CLI processes;
• bash scripts/check-abi-version.sh — snapshot, C header, TypeScript bindings, and ABI version are consistent;
• git diff --check.

Not run because this has no runtime, browser, libc, POSIX, VFS, or package-artifact changes:

• host Vitest;
• browser Playwright and ./run.sh browser;
• libc, POSIX, and Sortix suites;
• package rebuilds.

cargo fmt --check was attempted through the dev shell, but that shell does not currently provide cargo-fmt. The new Rust test follows the existing test style and compiled in both fork-instrument runs above.

[github-actions]

Phase B-1 matrix build status — pr-888-staging

ABI v18. 53 built, 0 failed, 53 total.

Package	Arch	Status	Sha
bc	wasm32	built	f45f740f
bzip2	wasm32	built	19695174
coreutils	wasm32	built	365366cf
curl	wasm32	built	2ae673c2
dash	wasm32	built	25bdf2cf
diffutils	wasm32	built	85e7d2a9
dinit	wasm32	built	177bfb90
fbdoom	wasm32	built	a61114ca
file	wasm32	built	09380c0b
findutils	wasm32	built	4a04839e
gawk	wasm32	built	05a3ab3f
git	wasm32	built	56ff3964
grep	wasm32	built	eb3c0049
gzip	wasm32	built	9d578800
kandelo-sdk	wasm32	built	fe4bd1c9
kernel	wasm32	built	355953e0
less	wasm32	built	4dbb52f7
lsof	wasm32	built	c8a88920
m4	wasm32	built	2910bb31
make	wasm32	built	ccf6bc63
mariadb	wasm32	built	518c5c54
mariadb	wasm64	built	4fa41b94
modeset	wasm32	built	b8d78740
msmtpd	wasm32	built	9ca2c667
nano	wasm32	built	ce3f0ed8
ncurses	wasm32	built	e5a0410b
netcat	wasm32	built	0994f5e6
nginx	wasm32	built	7882e5a9
php	wasm32	built	8d635623
posix-utils-lite	wasm32	built	4ce25d4e
ruby	wasm32	built	f70a1eea
sed	wasm32	built	80aa0020
tar	wasm32	built	dbcd5b7a
tcl	wasm32	built	55f6bfc5
unzip	wasm32	built	7102a849
userspace	wasm32	built	126d9bdd
vim	wasm32	built	d53d836a
wget	wasm32	built	b2ec19ac
xz	wasm32	built	41ce04bc
zip	wasm32	built	f69e2537
zstd	wasm32	built	d8e76ecf
bash	wasm32	built	7b9da8e5
mariadb-test	wasm32	built	8645154d
mariadb-vfs	wasm32	built	c199b6fb
mariadb-vfs	wasm64	built	2954778b
nethack	wasm32	built	dd04c9c8
vim-browser-bundle	wasm32	built	2c34f0fd
nethack-browser-bundle	wasm32	built	3e635d0c
rootfs	wasm32	built	8e7ae770
shell	wasm32	built	6b23d674
lamp	wasm32	built	04258e42
node-vfs	wasm32	built	07e99d2c
wordpress	wasm32	built	bd6dfe9d

_{Auto-generated; replaced on each push. Raw data in the publish-status workflow artifact.}

[brandonpayton]

Homebrew Sharutils exact consumer validation

I reproduced this PR's failure mode and fix against the identical linked, pre-instrumentation GNU Sharutils 4.15.2 shar input.

Input:

• 230,576 bytes
• SHA-256 43be9de13cc1e81d256d5d379fd7482c2524384dd349a2e14e291f59c6847466

Pre-fix instrumenter:

• 24 fresh processes produced 24 distinct outputs
• every output was 306,664 bytes
• two saved outputs differ at exactly 328 bytes: 164 synthetic local.set operands and their 164 matching local.get operands
• the variance comes from the regions.keys() walk that collects carryover plans before allocating per-subregion spill locals

Exact PR head f3c2c587c65ed7ed15507697efc2b381acd1a2d4:

• 24 fresh processes produced 24 byte-identical outputs
• output SHA-256 236b32fbe2122c1acf6e307bb17112f6083410fea312e4ed5c9ca0697e544e15
• wasm-validate --enable-all passed
• a reduced diagnostic that sorted only the carryover-plan region walk produced the same exact SHA as this PR

This independently confirms the randomized nested-region allocation cause and the canonical-order fix on a real Homebrew consumer. Sharutils bottle publication will remain gated on this PR and a post-merge rebuild. No ABI change is implicated by this evidence.

[brandonpayton]

Recording the pre-rewrite head before focusing this PR: f3c2c587c65ed7ed15507697efc2b381acd1a2d4 (currently stacked on fix/teardown-reclaim-readiness at bcd704a6aa7678f8ef326cb7f65003db971b8d02). I’m rebuilding the branch directly on current main, keeping only the fresh-process determinism regression, fixture, and current documentation. The production sorting change already landed through #907, so it will not be replayed.

[brandonpayton]

Closing this focused source PR because #934 now carries the unique fresh-process fork-output reproducibility fixture and current ABI-39 documentation at exact green head 260d75d5adc31b327f5591229dbfbb287d395fd6. The implementation already on main through #907 and stale ABI wording were deliberately excluded. The full staging and fork-instrument gates passed. This is not a claim that this branch was merged verbatim.