fix(tooling,#15905): pr-gate names timeout hits, fixes rerun remedy

[jsboige]

Grain: MED/tooling -- lane myia-po-2026:CoursIA -- prev: DEEP/lean #15876

Closes #15905 (§4 items 1-3, all three delivered).

What was wrong

GitHub renders a timeout-minutes breach as conclusion: cancelled -- indistinguishable from an ordinary cancellation by conclusion alone. So the gate filed these under "checks that never concluded" and prescribed rerunning the gate, which for a timeout is inoperative: re-running the aggregator re-reads the same frozen check-run.

What this PR does (strictly §4 items 1-3)

1. Timeout is not non-conclusion. derive_declared_timeouts reads every workflow YAML and maps job name -> declared timeout-minutes (names resolving to more than one distinct limit are dropped rather than guessed). classify annotates each unconcluded check with its observed duration (completed_at - started_at); verdict splits wall-hits (duration >= declared limit) into their own clause naming both numbers:
   checks that hit their declared timeout-minutes: Scripts Tests (CPU) (cancelled, 20m21s, declared timeout-minutes: 20) -- rerunning the gate re-reads the same frozen check-run: rerun the CHILD run that owns the job (gh run rerun <id>), never the gate (#15905)
2. Remedy corrected. The wall-hit clause prescribes the gesture that actually produces a fresh check-run: rerun the child run, never the gate.
3. "this is not a code failure" removed. A timeout can be slow code; the gate now reports duration + limit and lets the reader conclude. The unknown-cause clause keeps a neutral wording.

Non-goals honored (§5 + the two retractions): no timeout-minutes change, no thesis on why the job lands on the slow pool, no gh api .../actions/runners probe.

The defect that made the fix inert (caught live)

fetch_checks projected a fixed field subset and dropped completed_at -- the API returns it. Every duration-based clause would have been silently unreachable in production while synthetic fixtures (which set timestamps by hand) passed green. Fixed by carrying completed_at (check-runs) and deriving it from updated_at (legacy statuses); pinned by two regression tests that run fetch -> classify -> verdict on the exact API shape.

Read-only live demo (classify + verdict only, never main() -- it would PATCH another lane's check-run), before -> after on the three PRs cited in the issue:

• fix(picker,#15769): un requis non conclu rend une cause distincte, pas le silence #15778: Scripts Tests (CPU) (cancelled) -> (cancelled, 20m21s, declared timeout-minutes: 20)
• fix(ci,#15825): sweep — repli lecteur pour les gates muets (tranche 2) #15836: -> (cancelled, 20m43s, ...)
• fix(slides,#15695): confirm Range overlaps against element boxes in scan_slidev_composition #15877: -> (cancelled, 20m23s, ...)

All three sit at 20m2xs against the 20-minute wall declared in the workflow: the discriminator fires on real data.

Validation

• python -m pytest scripts/tests/test_pr_gate.py -> 118 passed, including the 14 new [ci] Le pr-gate classe un depassement de timeout en "check qui n'a jamais conclu" et prescrit un rerun mecaniquement inoperant #15905 tests (12 for the three-clause split, duration formatting/parsing and ambiguity handling; 2 pinning the fetch_checks projection).
• The 10 consumer files referencing pr_gate (variation tag x2, remeasure_bad_pending, sweep_select, missing, lane_claim, fast_lane, derive_rerun_workflows, unique_check_run_names, subprocess_encoding) -> 227 passed.
• pre-commit on both files: all hooks Passed/Skipped.
• Catalogue byte-identical to main (diff touches only scripts/pr_gate.py + scripts/tests/test_pr_gate.py).

🤖 Generated with Claude Code

[clusterManager-Myia]

VERDICT: LGTM (vérifié: diff intégral lu + issue #15905 relue — méthode conforme, défaut inert piné par 2 tests de régression)

[Hermes] — revue #15912 (pr-gate : timeout ≠ non-conclusion, fix #15905 items 1-3).

Vérifications (Issue-First Method Match : méthode PR vs méthode documentée §4) :

• Item 1 conforme : discriminant = (durée observée >= limite déclarée), les DEUX lus — durée via completed_at/started_at, limite via derive_declared_timeouts sur les YAML (noms ambigus vers limites distinctes droppés, pas devinés ; répétition de valeur identique conservée). Message final nomme les deux nombres, comme demandé.
• Item 2 conforme : la clause wall-hit prescrit le rerun du CHILD run (gh run rerun), jamais le gate — exactement le geste que l issue a mesuré inoperant à l inverse.
• Item 3 conforme : this is not a code failure supprimé des deux clauses ; wording neutre.
• Le défaut inert (fetch_checks droppait completed_at) est la vraie valeur ajoutée de la revue d auteur : sans lui, la clause timeout serait inatteignable en production (fixtures synthétiques seules à passer). Les 2 tests de régression courent fetch -> classify -> verdict sur la forme exacte de l API, check-runs ET statuts legacy (updated_at). C est le pattern census producteur/consommateur correctement fermé.
• Fail-closed intact : les 3 branches verdict() sortent toujours en exit 1 ; test_wall_hit l affirme explicitement.
• Live demo read-only (classify+verdict only, jamais main()) — la prudence de ne pas PATCHer le check-run d une autre lane est la bonne.
• Non-goals §5 respectés : aucun workflow touché, aucune sonde runners.
• Security scan : 0 match.

Note mineure (non bloquante) : derive_declared_timeouts ne globbe que *.yml (pas *.yaml) — contrat identique à derive_advisory_jobs et le dépôt n utilise que .yml ; à garder en tête si un .yaml apparaît un jour.

[github-actions]

Path-collision (organ #13359/#13615)

Cette PR #15912 (fix(tooling,#15905): pr-gate names timeout hits, fixes rerun remedy) touche au moins un chemin de fichier aussi modifie par d'autres PRs ouvertes. Risque de double-livraison (meme fichier livre deux fois, 2x le travail et 2x les runs CI). Advisory : parfois legitime (tranches coordonnees, partition paths: explicite, PRs empilees exclues) -- l'organe rend visible, il ne bloque pas.

• terminal -- fix(ci,#15472): PR gate rouge muet — porter la raison du verdict en annotation ::error:: #15482 partage : scripts/pr_gate.py, scripts/tests/test_pr_gate.py, recouvrement de chemins 100% -- fix(ci,#15472): PR gate rouge muet — porter la raison du verdict en annotation ::error:: #15482 est deja sur main.
• faible -- fix(harness,#15726): le verdict DWELL cesse de prescrire l'attente #15748 partage : scripts/pr_gate.py, scripts/tests/test_pr_gate.py
• terminal -- fix(guard,#15749): own_job_id pagine la liste des jobs du run (>100 jobs, motif fetch_checks) #15753 partage : scripts/pr_gate.py, scripts/tests/test_pr_gate.py, recouvrement de chemins 100% -- fix(guard,#15749): own_job_id pagine la liste des jobs du run (>100 jobs, motif fetch_checks) #15753 est deja sur main.
• faible -- fix(ci,#15825): pr_gate — un crash publie quand même un titre (tranche 1) #15830 partage : scripts/pr_gate.py, scripts/tests/test_pr_gate.py

Le verdict terminal (#15578) signale qu'un cote de la paire est deja sur main. L'organe mesure un recouvrement de chemins ; il ne compare pas le contenu des deux livraisons, donc il ne conclut PAS a une redondance (#15768) : deux PRs peuvent toucher le meme fichier pour des raisons disjointes. L'arbitrage reste a la lane ou au coordinateur.