[squid-protocol]

Part of #5137, #5142. No workflow edit.

Re-run measure (after #5141 / #5177 / #5190)

ci_rerun_stats.py mainframe-ground-truth.yml det-sweep.yml --since 2026-10-15 --timeout ground-truth=35 --logs

	before (10-13..14, 39 h)	after (10-15..16)
MGT re-run attempts	13	0
Det Sweep re-run attempts	9	1

The one Det Sweep re-run (run 38830516356 attempt 1, PR branch churn/5089) is registry_pull: shard db2 4/4 raised Db2ImagePullError: INFRASTRUCTURE (not a code defect): could not pull the pinned Db2 image icr.io/... after 4 attempts; det then failed with "its shard did not report". Truly external, already behind the bounded retry (#5097), so no new retry.

Change

tests/tools/ci_rerun_stats.py: new cause registry_pull, found with --logs (reads only the log of a job whose failed step is a Db2 one; marker PULL_MARKER = the Db2ImagePullError text). Without --logs, or with the log expired, it stays db2_start. Tests: marker in the failed job's log / in another job's / absent / no logs; a log never overrides a timeout or a non-Db2 step; _db2_logs fetches only failed Db2 jobs.

#5142 box 2: MGT "Gate against" step (2-4 min) - measured, no fix adopted

The step is the cold scan of the 3 keyed corpora, run serially by ground_truth_ledger.measure_all. I tried scanning them concurrently (process pool warm-up) and timed the full ground_truth_ledger.py check cold, pinned to 4 cores with GALAXYSCOPE_MAX_WORKERS=4 (the runner's size):

	run 1	run 2	run 3
serial (today)	87 s	95 s	86 s
3 scans concurrent	70 s	74 s	84 s

About 15% and inside the noise: one galaxyscope already uses the 4 cores (user time 3m36 vs wall 1m27 serial), so the corpora do not overlap usefully. Below my 20% bar, so not shipped. The cost is the engine scanning ~3 corpora; a real cut needs a warm scan cache keyed on the files that change a scan (the engine HEAD key makes every engine PR cold by design).

Box 1 (dsf unchanged-check) was fixed by #5190 (content-hash key); its cache is cold until main's next push touching the generation paths.

Sibling gaps

• MGT gate step 90-250 s cold scan: measured above, not reduced; needs a scan-cache key narrower than the engine HEAD (changes what a stale DB can mean for a gate; owner call).
• Det Sweep re-runs: 1 sample in the window; re-measure after a week.

Gates

pytest tests/tools/test_ci_rerun_stats.py + test_ci_retry.py 33 passed (plain pytest from /tmp); pr_gates.py --fast 6/6. tests/tools only: no golden, ledger or ratchet effect.

Lessons for agents

• gh api .../jobs/ID/logs refuses output with terminal escape sequences: pass --allow-escape-sequences.
• A process pool, not threads, for concurrent mainframe_corpus.scan: it sets and restores os.environ["PYTHONPATH"] around its subprocess.

🤖 Generated with Claude Code

[squid-protocol]

No description provided.