Three changes to make `pnpm test` reliably minimal and fail gracefully:
- @fusion/core is now a memory-envelope/wide-fan-out package (was unguarded).
It's the hub nearly everything imports (~354 test files), so a core source
edit made `vitest --changed` expand to ~the whole core suite and blow past the
engine's 15-min verification kill -> SIGKILL + task restart. Adding it to
SCOPED_AFFECTED_MEMORY_ENVELOPES applies the wide-fan-out guard (run only
directly-changed core tests, else delegate) and the bounded env. core is NOT
gate-covered, so delegation warns loudly rather than false-greens.
- Lower CLASS_BUDGET_BANDS.changed ceiling 20min -> 13min so the script watchdog
fails a runaway local lane itself (exit 124, no restart) BEFORE the engine's
15-min kill restarts the whole task. A tightening, not a timeout-widening.
Guard test pins ceiling < 900_000ms.
- Raise scoped-affected worker fan-out 1 -> 4 (operator decision). Was 1 only
for OOM safety (FN-6854/FN-6874); the fan-out guard now bounds the set so the
hundreds-of-files OOM driver no longer reaches these workers. Heap stays
6144MB/worker (~4x6GB on the lane) — revisit if a RAM-constrained CI runner
OOMs. Trades FN-5048 worker-knob guidance for throughput, scoped to the
bounded affected lanes only.
Tests: test-changed 117/117, watchdog 15/15, eslint clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- Reformat shard-floor justification as an FNXC:TestInfrastructure comment
(project-standards: AGENTS.md FNXC_LOG convention).
- Clarify that the shard and dashboard-lane 15min floors are not coupled and
may diverge (maintainability: avoid implying an unenforced contract).
- Add a regression-guard test pinning shard.floor=15min and asserting a 525s
derived budget clamps up to the floor, so an accidental revert to the old
5min floor fails loudly (correctness + testing + project-standards).
The Full Suite (non-blocking) workflow has been red for 30+ runs on main.
Diagnosis: the @fusion/engine [1/2], [2/2] and @fusion/core [2/2] shard
slices were SIGKILLed at their watchdog budgets (405s/405s/338s), not because
they hang but because those budgets are too tight for current wall-clock.
Local baselines (this machine, all pass, exit 0):
- engine [2/2]: 145s wall / 309 files
- core [2/2]: 283s wall / 172 files (old budget was only 338s!)
deriveBudgetMs tightens the budget to expected*3.5 whenever the committed
scripts/test-timings.json is <30d old. The snapshot (2026-06-03) undercounts
the import- and real-git-subprocess overhead of these heavy slices, so the
'fresh' snapshot produced a too-tight, false-kill budget on slower CI runners
-- the exact failure mode the floor/ceiling band exists to prevent.
Fix (plan KTD-2): raise the shard band floor 5min -> 15min so the heaviest
slices can't be tightened into a false-kill, while a true hang is still bounded
far under the job's 60min ceiling. Mirrors the dashboard-lane heavy-lane floor.
Follow-up: refresh scripts/test-timings.json from a default-branch CI run.
- watchdog: escalate forwarded SIGINT/SIGTERM/SIGHUP to SIGKILL after grace so
external cancellation can't hang for the full budget (coderabbit major)
- watchdog: route onProcExit through signalGroup for injection consistency (greptile)
- watchdog: add cwd option; test-changed passes rootDir so pnpm runs from repo
root regardless of invocation cwd (coderabbit major — preserved original run() cwd)
- dashboard runner: validate/clamp FUSION_RUN_VITEST_* env so a malformed value
can't NaN-disable the watchdog (coderabbit)
- tests: verify exit-listener cleanup, forwarded-signal escalation, cwd passthrough
- plan doc: per-class-ceiling fallback wording (not median); label Output Structure fence
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
U1/U2/KTD-6: extract the dashboard heap-runner's process-group kill lifecycle
into a shared scripts/lib/run-vitest-watchdog.mjs with per-class budget bands
(timings only tighten within a generous ceiling) and an inline hang-diagnostics
snapshot. Delegate run-vitest-with-heap.mjs to it (behavior preserved). Add
timeout-minutes backstops to all full-suite.yml test jobs so a wedged run can no
longer hang to GitHub's 6h ceiling.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
