fix: refresh clownfish dashboard aggregates

[brokemac79]

Summary

• add a daily aggregate publish pass so the README dashboard refreshes even when there are no new cluster artifacts to ingest
• let aggregate publishes continue past empty backfill/review-filter outputs while preserving the normal no-op path for the frequent poller
• record exact cluster-batch child job metadata, including per-matrix conclusion/status and selected run attempt, so one failed child no longer stamps every child from the parent workflow conclusion
• keep review filtering and Windows artifact-path identity inference aligned with exact child run ids

Root Cause

The scheduled publisher could exit before running the aggregate/dashboard path when no missing cluster artifacts were found. The existing frequent schedule also passed --skip-aggregate, so successful runs could leave the README dashboard stuck on an old timestamp.

The failed-cluster count was also inflated for cluster-batch.yml artifacts because backfill wrote one parent workflow conclusion for every matrix child. A single failed matrix job could make all sibling child records look failed.

Validation

• node --check scripts/backfill-cluster-results.mjs
• node --check scripts/filter-reviewed-artifacts.mjs
• node --check scripts/result-run-identity.mjs
• node --test test/publisher-artifact-recovery.test.mjs test/publish-result.test.mjs test/result-run-identity.test.mjs (14 pass)
• Codex review before PR: accepted one P2 finding about rerun attempts, fixed it, reran focused tests, then reran codex review -c service_tier='"fast"' --uncommitted; final review reported no actionable bugs.

Notes

• Full npm test was attempted during this investigation but is not clean on this Windows checkout because of existing fixture/environment issues unrelated to this patch, including fake .mjs command execution on Windows, GitHub 404s in preflight fixtures, and malformed archived job fixtures.