The shared ratchet now rejects compiler/setup failures before filtering app paths or updating a baseline. This closes the pathless-error false green found during the screenshot follow-up. No application, slide, narration-input, dependency-version or baseline change is included.
Behavior and evidence#
- Global TypeScript diagnostics, unsupported exits, spawn/signal failures and unexplained nonzero exits fail. stdout/stderr continuations remain separate. Located dependency errors and known app-owned fingerprints/counts retain their existing boundary; new app-owned errors still fail.
- Recognized npm warnings remain visible but cannot justify a failed exit by themselves. Compiler color settings are normalized. Fatal output is flushed before exit, including the final error after a large dependency-error dump.
- The cached wrapper hashes both gate implementations, including uncommitted
edits. Nx's default typecheck inputs include the tool's
.mjsfiles. A local disposable-checkout test exercises reuse, gate/wrapper invalidation and failure non-caching. This is not hermetic dependency/environment cache validation. - All 40 Node checks pass, including six installed-TypeScript integration
cases. The identical test/fixture bytes against the committed
d5b9a56694gate, wrapper and Nx configuration produce 24 failing / 16 passing cases. That deliberately failing negative control is not counted as a passing gate. - All three targeted uncached Nx typechecks pass: OSHUN web, BFF and prompt-injection evaluations. Web/BFF report zero app-owned backlog; their transitive dependency diagnostics are not asserted clean. The direct strict page-inspection test compiler check also passes outside the ratchet.
- Scoped root ESLint and Prettier pass. Local YAML parsing verifies both CI path triggers and the serial test step after setup/before affected selection, without conditional bypass or continue-on-error. OSHUN CI now runs the suite. This record does not claim a hosted Actions run; actionlint was not installed.
The first real consumer attempt rejected launcher warnings and Nx stopped after
the BFF failure; web/evaluations were not run in that attempt. A BFF-directory
probe reproduces the inherited NO_COLOR/FORCE_COLOR warning. A separate root
npx probe also exposed warnings about pnpm-only settings. Both warning classes
informed the correction; the failed compiler dump was truncated by immediate
exit, so it is not treated as a complete stderr diagnosis. The final 4,000-error
regression verifies complete output through its last stderr marker.
Initial test setup also needed correction: the disposable cache had to be
excluded from its own Git key; an unmatched include reproduces global no-input
errors whereas files: [] emits a located config error; large test output must
be generated in the child instead of exceeding Linux's environment-string
limit. Final before/after runs use the corrected, identical fixtures.
Reproduction and scope#
node --test --test-concurrency=1 tools/typecheck/ratchet.test.mjs
NX_DAEMON=false NX_ISOLATE_PLUGINS=false NX_SKIP_NX_CACHE=true \
python3 tools/presentations/run-supervised.py node_modules/.bin/nx run-many \
--target=typecheck --projects=@oshun/web,@oshun/bff,@oshun/prompt-injection-evals \
--parallel=1 --nxBail --skip-nx-cache
pnpm --dir apps/oshun/web exec tsc --noEmit --pretty false -p tsconfig.page-inspection-tests.json
python3 docs/presentations/presentation-center/verification/typecheck-integrity-2026-09-07/verify.py
python3 docs/presentations/presentation-center/verification/typecheck-integrity-2026-09-07/test_verify.py
The host has 15 GiB RAM and 31 GiB swap. Expensive gates ran serially with a memory supervisor and a 20-second external process/resource monitor. The final Nx run took 3:43; maximum child RSS was 7,665,580 KiB and the external monitor's lowest rounded available-memory sample was 3.5 GiB. No OOM or threshold stop occurred. Other workspace processes were left alone. The live narrator continued; no presentation content or partially rebuilt duration catalog was published.
state.json binds named inputs and retained evidence, not every transitive file.
The four receipt tests check valid evidence plus missing/changed inputs and
altered evidence. Full log copies strip terminal escapes/trailing whitespace
only; the initial consumer failure is an explicitly labeled excerpt with the
raw log's size/hash. Resource JSON retains memory samples, not other agents'
process listings.
This is not an all-consumer sweep, production build, fresh browser/service test
matrix, full Eve release or completion of the Presentation Center checklist.
The prior screenshot receipt is intentionally unchanged. Its hash gate now rejects the superseded ratchet input, correctly identifying it as historical; its browser/service results have not been rerun or relabeled as new evidence.
Separate narration continuity correction#
The read-only baseline audit compares the pinned manifest at
79814f43c6a7271747dd652ebbaa15ac54c2659a with the unchanged live speech-script
fingerprints. Exactly 177 tracks were current, and none of the eight pending
agentic replacement IDs belonged to that set, including agentic-studio-failure.
This supersedes the contrary baseline sentence in the older recovery draft and
receipt; do not introduce an exception to the original-177 preservation audit.
It does not apply the drafts or verify new audio decoding, pacing or listening.
python3 docs/presentations/presentation-center/verification/typecheck-integrity-2026-09-07/audit_narration_baseline.py