Cross-cutting review polish:
- worker.go: id/type → job_id/job_type; one stray attempts → attempt for
consistency with the other 4 sites in the same file.
- runners/workspace_fetch.go: ws_id → workspace_id (matches the dispatched
notify.Message metadata key).
- runners/worktree_prune.go: id → worktree_id (same reason).
- runners/worktree_prune_test.go: add TestWorktreePrune_NoRowsIsNoOp for
symmetry with the no-op tests in the other 4 runner test files.
- migrations/0005_jobs.down.sql: comment that DROP TABLE jobs cascades to
drop the three idx_jobs_* indexes automatically.
Module.Stop used time.After which leaks the timer if the done channel
wins first; switch to time.NewTimer + defer Stop() to align with the
codebase convention.
scheduleEntry silently returned {0, false} for unknown runner names —
typos in production wiring were invisible. Return (entry, known) and
log Warn at NewModule registration time so misconfiguration surfaces.
Previously, write-back errors from Complete/DeadLetter/FailWithRetry
were silently dropped via underscore assignment. A transient DB
failure here would leave a job stuck (still 'running' with active
lease) until the reaper resets it, with zero observability into why.
Log each write-back error at Error level so operators can correlate
stuck jobs with DB symptoms. Code quality review of T8.
q792602257