Filed by the spec seat on the maintainer's instruction (2026-08-20 live chat, verbatim: 「立卡」, accepting layer ② of the queue-incident remediation menu — the maintainer thereby also approved the nightly CI-minutes spend this card exists to flag). Provenance: #10115 · PR #10120 · #10126 (layer ①).
The blind spot
PR-side CI runs the affected subset; the merge queue runs the FULL suite under shard parallelism on 2-core runners. Between them there is no rehearsal signal: a defect that only expresses under full-suite load (tonight: a cold-transform cost inside a 5000ms clocked window, green on its own PR at 01:53Z) stays invisible until it detonates inside the queue — where every red ejects an innocent PR and rebuilds everything behind it (30 failed builds / 24h, main frozen 2h+).
The change
Add a schedule: trigger (nightly) to the existing full-suite workflow path — the same shard definitions the merge queue uses, run against main, decoupled from any PR. A red names the offending test directly, at a time when it burns nobody's PR. Wire its failure into whatever notification surface the repo already uses for main-branch reds (no new channel).
Leads, not spec: reuse ci.yml's merge_group job matrix under a schedule event guard; mind the required-context name pin (check:required-context family) so job names do not drift; the run should be skippable in forks.
Cost note, explicit: one full-suite run per night of CI minutes — approved by the maintainer's 「立卡」 on the menu that stated this cost.
Dedup: no existing card proposes a scheduled full-suite rehearsal (searched: schedule / nightly / full suite / merge_group rehearsal).
Suggested size/model: S, opus (workflow edits carry name-pin traps).
Filed by the spec seat on the maintainer's instruction (2026-08-20 live chat, verbatim: 「立卡」, accepting layer ② of the queue-incident remediation menu — the maintainer thereby also approved the nightly CI-minutes spend this card exists to flag). Provenance: #10115 · PR #10120 · #10126 (layer ①).
The blind spot
PR-side CI runs the affected subset; the merge queue runs the FULL suite under shard parallelism on 2-core runners. Between them there is no rehearsal signal: a defect that only expresses under full-suite load (tonight: a cold-transform cost inside a 5000ms clocked window, green on its own PR at 01:53Z) stays invisible until it detonates inside the queue — where every red ejects an innocent PR and rebuilds everything behind it (30 failed builds / 24h, main frozen 2h+).
The change
Add a
schedule:trigger (nightly) to the existing full-suite workflow path — the same shard definitions the merge queue uses, run againstmain, decoupled from any PR. A red names the offending test directly, at a time when it burns nobody's PR. Wire its failure into whatever notification surface the repo already uses for main-branch reds (no new channel).Leads, not spec: reuse
ci.yml's merge_group job matrix under ascheduleevent guard; mind the required-context name pin (check:required-contextfamily) so job names do not drift; the run should be skippable in forks.Cost note, explicit: one full-suite run per night of CI minutes — approved by the maintainer's 「立卡」 on the menu that stated this cost.
Dedup: no existing card proposes a scheduled full-suite rehearsal (searched: schedule / nightly / full suite / merge_group rehearsal).
Suggested size/model: S,
opus(workflow edits carry name-pin traps).