Promotion 2026-09-23: staging → main

[modelmirror]

Promotion batch — 2026-09-23

Staging → main, ahead of the 2026-09-28 long conference. No process digest moves: no prompt, predictor/evaluator config, DEFAULT_MODELS, ENGINE_RETRIEVAL or MCP manifest entry is in the batch, and tests/test_process_version.py -k blessed is green at the staging head. No forward cell's execution path changes: live run-predict / run-evaluate call their engines through the actions and CLIs directly. The shared pipeline/runner.py change (#1959) reaches only local-cascade (engine-smoke) and the back-test.

PR	What	Class
#1958	Weekly digest restructured around three production windows	ops reporting; run-ops.yml comments
#1959	Back-test: a failed engine call is one lost cell; a spent quota stops that engine, never the campaign	back-test; shared runner retry classification
#1962	qp-topic labeler offers claude-opus-5-5; Claude CLI pin 2.1.259 → 2.1.280 (analytics, integration-test, back-test)	workflow
#1970	boto3/botocore, pypdf 6.19.0, ruff patch bumps	deps (lock only)
#1972	Back-test per-cell leakage disclosure record (recorded, not excluded)	back-test report schema; metrics/README reading rule
#1973	Plain-language case summaries lane (summarize.yml, Claude Sonnet 5, environment's Anthropic key)	new workflow; ci.yml main-base admits summaries/refresh; SECURITY.md
#1974	Back-test runs each engine's replay cells in its own lane	back-test; run-backtest.yml comments

Held out on purpose (drafts): #1964 (Opus 5.5 predict/evaluate, proc-v9 — after the long-conference release), #1971 (actions group incl. claude-code-action → v1.0.231, which bundles CLI 2.1.278 < 2.1.280), #1844 (engine-smoke provisioning — conflicts with this batch; rebase after).

Engine evidence: a full green scenario=all run at 98c4607c1 (run 35901466753; all six engine-smoke / engine-actions-smoke legs green — this matters because #1962 moves the Claude CLI the engine-smoke leg runs). The head then moved by a data-only re-sync (seed commit 81a0153aa, corpus/corpus.db.ref only), which cannot reach a cell, so the head is carried by an all-offline run plus the promote:skip-engine-smoke label, per When the head moves after a green all in docs/pipeline.md. The engine evidence for this batch's code is the 98c4607c1 run above.

Effect checks (report each as it lands)

1. Summaries lane registered, held behind the rehearsal. summarize.yml appears in Actions once this merges:

   gh workflow list --all | grep -i summar
   

   Then the staging rehearsal — the lane's first real model call:

   gh workflow run summarize.yml --ref staging -f limit=10
   

   Expected: plan reports ~10 planned with a cost range; the run waits on review; after release generate reports ~10 written with cost from response usage (≈$1–2 total), files in the case-summaries artifact, no PR. The summaries: plain-language case summaries per snapshot #1968 spot check (~10 summaries against their filings, factual errors and slant) is done on that artifact.
   Until the spot check passes, reject the daily main run's review hold (03:43 UTC) — it will plan the full ~195-case backfill. Once released after the spot check, it opens a summaries/refresh PR with ~195 files:

   gh run list --workflow summarize.yml --branch main --limit 3
   gh pr list --head summaries/refresh --state open --json number,title,url
   

2. Weekly digest (feat(ops): restructure the weekly digest around three production windows #1958). The Monday 2026-09-28 digest opens with ## Produced this week, then ## Produced this month and ## Produced this term (OT2025, …) before ## Backtest results, with no ## Health questions or ## Analytics state.
3. Back-test lanes (perf(backtest): run each engine's replay cells in its own lane #1974) and disclosure record (feat(backtest): per-cell leakage disclosure record #1972, fix(backtest): a failed engine call is one lost cell, and a spent quota stops that engine — never the campaign #1959). On the 2026-10-03 run-backtest cron (spread, once released):

   gh run list --workflow run-backtest.yml --limit 5
   gh run view <run-id> --json jobs --jq '.jobs[] | select(.name == "backtest") | ((.completedAt | fromdateiso8601) - (.startedAt | fromdateiso8601)) / 60'
   gh run view <run-id> --json jobs --jq '.jobs[] | select(.name == "backtest") | .databaseId'
   gh run view <run-id> --log --job <backtest-job-id> | grep -oE '\[(claude-code|codex|gemini)\] ' | sort | uniq -c
   

   Expected: all three engine prefixes interleaved in the log; the backtest job well under the 330-minute ceiling (~40–60 min of cells at 10 petitions); no ::error::cert-backtest annotation and no harness-error in provenance.lost_cells; the report carries provenance.disclosures and provenance.disclosure_tally. On a day gemini's allowance is spent, gemini's cells read quota-exhausted in lost_cells and the job log prints the quota drop line, with the campaign completing.
4. Opus labeler option (ci(qp-topic-label): offer claude-opus-5-5; claude CLI pin → 2.1.280 #1962). On a prod qp-topic-label dispatch with label_model=claude-opus-5-5, the labels artifact's batch records labeler: claude-code/claude-opus-5-5. (Rehearse on staging first if wanted.)
5. Deps (chore(deps): Bump the python group with 3 updates #1970). No run-time observable beyond a green gate; pypdf's extraction change is limited to Arabic-Indic digits.

🤖 Generated with Claude Code

[modelmirror]

Tagged promotion/2026-09-23 (annotated) on merge commit 32e0804. Effect check 1, first half: gh workflow list --all | grep -i summar → summarize active 365433562 — the lane is registered. The staging rehearsal and the other checks are pending; each is reported here as it lands.