Backfill project_id on first_run test_case_events rows left at the zero sentinel

Context

test_case_events.project_id (Int64 DEFAULT 0, denormalized in migration 20260721115000) is now relied on for project-scoped reads: PR #12215 scopes the quarantined-tests aggregate on it. While validating that PR we confirmed every state-change event type (muted, skipped, unmuted, unskipped, marked_flaky, unmarked_flaky) carries a real project_id — but first_run events do not: a subset of them still sit at the project_id = 0 sentinel.

How the zeros got there

Two populations, neither ongoing:

1. Orphans predating the backfill. The 20260721115000 mutation resolved project_id through a dictionary built from test_cases; first_run rows whose test case had since been deleted resolved to the 0 default and were left as-is.
2. A post-backfill write window. create_first_run_events has set project_id: run.project_id since #11977, but zero-rows continued for roughly a week after the backfill before stopping abruptly — consistent with a lagging ingestion component (e.g. the separately-shipped xcresult-processor image) still running a pre-#11977 release. A portion of these rows belong to test cases that still exist, so they are cleanly fixable.

No zero-rows have been written since; the current write path is correct.

Impact

None today: first_run is an audit-log-only event type, and its reads are scoped by test_case_id. But the column now looks uniformly trustworthy — every other event type is fully backfilled — so the next project-scoped query over events (exactly what PR #12215 added for muted/skipped) would silently miss these rows. This is a data-hygiene trap worth closing while the context is fresh.

Proposed fix

1. Re-run the dictionary backfill mutation from 20260721115000 (same advisory-lock pattern) for event_type = 'first_run' AND project_id = 0 — this fixes the post-backfill window rows whose test cases exist.
2. Decide on the true orphans (test case no longer in test_cases): either delete them (they can never render in any UI that starts from a test case) or document that project_id = 0 on first_run means “orphaned audit row”.

Out of scope

PR #12215 is unaffected: its scoped subquery reads only muted/skipped events (fully backfilled, zero mismatches against the owning test case’s project), and the flaky-tests subquery scopes by test_case_id.

🤖 Generated with Claude Code

No GitHub comments yet.