* test(#2650): failing-first coverage for the mis-sized stall-watch bound
Lands the regression matrix before the fix. Two assertions fail deterministically on this commit because the helper still calls raw spawnSync with a hard-coded 10000ms bound:
1. The process seam is never reached, so a mock.method spy on it records nothing and the class-norm bound cannot be observed.
2. A raw spawnSync result carries no outcome/timedOut field at all — which is precisely why the windows shard 1/3 failure printed 'null !== 0' instead of naming the timeout or the bound it exceeded.
The bound is sized for the wrong class of call. extractStallHelpersBash slices the entire bash fence, whose runtime-launcher preamble resolves gsd-tools.cjs and really runs two config-get lines — two Node spawns, measured 236ms against 2ms for the fallback the existing comment claims fires (118x). tests/helpers/timeouts.cjs already owns this class as HOOK_FANOUT_TIMEOUT_MS and its comment records the identical failure on PR #3285.
Refs #2650
* test(#2650): bound the stall-watch fan-out with the class norm and report it typed
Drives the failing-first coverage from 15aaef4a7 green (5 failures -> 0). runBashScript now routes through tests/helpers/process-seam.cjs instead of a hand-rolled spawnSync, which CONTRIBUTING.md already requires of anything that shells out.
The bound moves from a hard-coded 10000ms literal to HOOK_FANOUT_TIMEOUT_MS. That constant already exists for exactly this shape — a bash invocation that fans out to nested subprocesses — and its own comment records the identical failure on PR #3285: a bound sized for the wrong class, not a slow machine. This script is that shape: the extracted fence opens with the runtime-launcher preamble, which resolves gsd-tools.cjs and really runs two config-get lines.
The result is now typed. spawnSync reports a kill as status:null, so an exceeded bound reached the call sites as 'null !== 0' — naming neither the timeout nor the bound it exceeded. OUTCOME.TIMED_OUT names itself. status is still aliased from exitCode so the five existing assertions read unchanged.
Corrects the comment that caused the mis-sizing: it claimed gsd_run is undefined so the '|| echo' fallback fires. It is not — the preamble defines it, and the two Node spawns are real (236ms vs 2ms, 118x). They are deliberately left in place; removing them would change what the extracted script executes.
Also drops two now-dead '{ timeout: 10000 }' call-site options. The seam reads only its own documented keys, so those would have been silently ignored while still reading as a 10s bound.
Refs #2650
* test(#2650): make the boundary test a real value-domain boundary
Standards review flagged that the previous 'bound boundary' test repeated the TIMED_OUT arm the test above it already covers, and was not a limit-1/limit/limit+1 in any meaningful sense — an exact-millisecond timing edge would have been a race, not a boundary.
Replaced with a boundary on the VALUE DOMAIN of the bound itself: 0 and -1 must be rejected with TypeError, and 1 (the smallest positive value) must be accepted. Zero is the load-bearing case — spawnSync reads it as 'no timeout at all', which is exactly the unbounded-spawn hazard local/no-unbounded-spawn exists to prevent, so the seam rejects it rather than honouring it.
Also asserts the rejection path does not leak the script temp dir, since the throw escapes through runBashScript's finally. Soak: zero=TypeError, negative=TypeError, one=no-throw, leaked dirs=0.
Refs #2650
---------
Co-authored-by: sim <sim@local>
Folds 15 tests/fix-*.test.cjs regression files (191 test() blocks) into
their module's main suite, per the wave decomposition of #3315 (H3 of
epic #3053). 187 blocks land in 8 existing suites (4 exact-duplicate
cases dropped, documented inline); 4 blocks move via git mv into 2 new
suite files with no prior coverage to merge into. Zero production
behavior change.
Also tightens two H1 (#3313) ratchets that the fold's own file-count
reduction moved past their grace window, per the ratchets' documented
dual failure mode (a stale/too-loose baseline fails exactly like a
novel violation):
- lint-test-file-count.allowlist.json: removes the stale "audit" entry
(folding fix-2766 into tests/uat.test.cjs drops that module back to
its 2-file cap).
- lint-allow-test-rule-refs.ceiling.json: lowers maxFiles 314 -> 309,
the real post-fold high-water mark (gsd-test's own repo-baseline
test caught this — CI, not a human, found it).
Two orthogonal review passes (Standards+Spec code-review, isolated
security-review) found and this commit fixes two issues before push:
a genuinely-distinct #2287 test case (file-absent vs. file-present-
resolved) that a prior fold pass had wrongly dropped as a duplicate —
restored verbatim into tests/uat.test.cjs; and a missing same-line
allow-test-rule citation on the #2196 block in
tests/debug-session-management.test.cjs, added for consistency with
its sibling #2257 block.
lint-removed-but-needed also caught two stale doc references to the
now-folded-away fix-2285-claude-orchestration-wiring.test.cjs filename
(docs/adr/1143-claude-orchestration-capability.md,
gsd-core/references/execute-phase-response-language.md) — updated both
to point at tests/claude-orchestration.test.cjs, its new home.
Co-authored-by: sim <sim@local>
