@@ -164,39 +164,15 @@ two-client same-device test pins both clients to one device via `pinnedDevice` s
164164different devices, driving that contention deterministically.
165165
166166Every failure prints the offending seed and the exact ` TORTURE_SEED=<n> pnpm test:concurrency-torture `
167- replay command. The PR gate runs the fast default sweep through the Node integration lane
168- (` test:integration:node ` ); the ` Concurrency Torture Nightly ` workflow sweeps a much larger seed range
169- on schedule and, per #1430 , emits a machine-readable envelope (schema version, commit SHA, tool/config
170- hash, seed range, duration, result) via ` TORTURE_ENVELOPE=<path> ` , uploaded as the
171- ` concurrency-torture-envelope ` artifact. The envelope is written once, after ** all** lane tests
172- settle, and reports ` fail ` if any of them (sweep, replay self-check, or forced-contention guardrail)
173- failed — a later-failing guardrail can never be published as a passing envelope.
174- Optional knobs: ` TORTURE_CLIENTS ` , ` TORTURE_OPS ` .
175-
176- ### Scheduled-lane health watcher (#1430 )
177-
178- Scheduled lanes can fail or stop running for weeks while PR CI stays green, so the `Scheduled Lane
179- Health` workflow ( ` .github/workflows/scheduled-lane-health.yml`) watches the watchers. It discovers
180- every ` schedule: ` -triggered workflow from ` .github/workflows/ ` (the list is ** derived** , not
181- hand-maintained), reads each lane's recent scheduled runs via the GitHub API, and opens/pings a single
182- tracking issue when a lane has not ** succeeded** within two of its own cadences — which covers both a
183- lane gone dark (no runs) and one failing every cadence. Freshness is measured from an anchor: the last
184- successful run, or — when a lane has never succeeded — when its ** schedule became active** : the
185- committer/landing time of the most recent * unscheduled→scheduled* transition of the workflow on the
186- default branch's first-parent history, derived from git (` run.ts ` walks the file's history and parses
187- each revision's YAML, so it uses landing time not author time, ignores a ` schedule: ` mentioned only in
188- a comment, and survives a schedule removed and later re-added). The workflow needs full history, so the
189- watcher checks out with ` fetch-depth: 0 ` . This is deliberately ** not** the workflow's ` created_at ` ,
190- which predates a schedule added later to an old workflow and would spend the grace before the lane was
191- ever scheduled. When git history is unavailable (e.g. a shallow checkout) it falls back to the earliest
192- scheduled run, then to the current time. Anchoring on schedule-activation gives a ** newborn lane its
193- grace** : a lane whose schedule is younger than two cadences (zero runs yet, or a single failed first
194- cadence) is not alerted, because two cadences cannot have been missed/failed yet. The cadence is
195- estimated per workflow from its cron expression. The pure verdict is unit-tested in
196- ` scripts/scheduled-lane-health/model.ts ` and the git-history derivation in ` run.ts ` (both gated on PRs
197- via ` model.test.ts ` + ` run.test.ts ` ); the GitHub API I/O and issue open/ping run nightly. New
198- scheduled lanes need no wiring here — emitting the standard envelope and carrying a ` schedule: ` trigger
199- is enough to be watched.
167+ replay command. The lane lives under ` test/integration/nightly/ ` , deliberately ** out** of the
168+ ` test:integration:node ` glob so it is not an accidental PR-time run: the PR gate runs a fast default
169+ sweep via an explicit ` Run seeded concurrency torture lane ` step in the Integration job, and the
170+ ` Concurrency Torture Nightly ` workflow sweeps a much larger seed range on schedule. The nightly run
171+ emits a machine-readable envelope (schema version, commit SHA, tool/config hash, seed range, duration,
172+ result) via ` TORTURE_ENVELOPE=<path> ` , uploaded as the ` concurrency-torture-envelope ` artifact. The
173+ envelope is written once, after ** all** lane tests settle, and reports ` fail ` if any of them (sweep,
174+ replay self-check, or forced-contention guardrail) failed — a later-failing guardrail can never be
175+ published as a passing envelope. Optional knobs: ` TORTURE_CLIENTS ` , ` TORTURE_OPS ` .
200176
201177## Speed rules (experiment-backed, 2026-07-04)
202178
0 commit comments