[zhaozhaozz]

An older workflow run can pass after a rerun while its original creation time stays outside incremental polling. Refresh notified run IDs during the existing collection instead of adding another Actions-history scan or dispatch for each completion.

• Fetch at most 20 nonsecret rerun hints from the configured bot with the existing collector identity. Prioritize direct run/attempt and necessary report reads before bulk backfill, within the current request, time and artifact budgets.
• Keep the existing latest-attempt CI projection, report caches and historical attempts. Generate a receipt only after progress and public data are atomically committed; use a fresh OIDC identity to acknowledge the successful subset. Lookup, detail, budget or publication failures and lost acknowledgements leave hints pending.
• Keep schedules, workflows, debounce, OIDC trust, token permissions and E2E record versions unchanged. An older/unavailable hint endpoint warns and leaves ordinary collection running.

Companion bot: ScienceDiscovery/sciencediscovery_bot#23. The endpoint can be deployed first; unavailable endpoints leave ordinary collection running and pending hints remain at the bot.

Validation at 88b87055a24688a720f3558fd490193de65ec381:

Type	Command / scenarios	Result
GitHub Actions	Dashboard checks on this head	PASS, verify in 1m12s
Regression tests	PYTHONDONTWRITEBYTECODE=1 TMPDIR="$PWD/.tmp" python3 -m unittest discover -s tests	PASS, 165 tests
CLI journey	PYTHONDONTWRITEBYTECODE=1 python3 test/journey-rerun-collection.py; out-of-window failed run becomes green with its latest attempt and an atomic publication receipt	PASS
Browser regression	node .e2e/node_modules/playwright/cli.js test --config test/playwright.config.cjs; all dashboard journeys, including Real E2E marker geometry, carried/missing-score dashed lines and first-measurement boundaries	PASS, 27 journeys
CLI recovery	Same command; source lookup failure preserves the old conclusion; publication conflict creates no acknowledgement	PASS, 3 journeys in total

The journeys run the real collector/publisher process against a local HTTP stand-in and check the Git publication and post-publication receipt. No live collection was triggered.

The existing Real E2E curve journey still expected the pre-October-15 .score-failed marker, a zero before the first measurement and a broken line across missing scores. Main already renders .score-x markers and dashed bridges after the first measured score. Update the regression assertions and documentation to this existing behavior, including marker coordinates, bridge geometry and missing-score explanations; production chart code and fixtures remain unchanged.