Problem
The installed reporting production gates run close to their hard deadlines and provide little or no progress output while the isolated child process is active. This makes a production deadlock, a slow phase, and a runner-capacity problem look alike until the outer timeout expires.
The rc.4 adoption incident in #1195 is the immediate example:
vcs-pg-installed-production produced no stdout/stderr for 1,800 seconds- the exact local equivalent completed in 1,416.56 seconds
- CI learned only that the child exceeded the aggregate deadline
- useful PostgreSQL evidence existed, but it was not correlated to a named inner phase or wait state
This issue is for test performance and observability. It must not replace or conceal the production concurrency investigation in #1195.
Desired changes
- Emit bounded, structured progress for each production-contract phase.
- Apply per-phase deadlines beneath the aggregate job deadline.
- Capture process-tree and PostgreSQL wait-state metadata when progress stalls.
- Identify which test/seed/replay/restart phase is active without printing fixtures, provider responses, credentials, or customer data.
- Profile duplicated setup and expensive work across VCS/sdist and base/PG cells.
- Safely reuse or shard immutable build inputs where isolation guarantees remain intact.
- Retain original logs and cleanup outcomes as CI artifacts.
Acceptance criteria
- A stalled lane reports its active phase and last progress timestamp before termination.
- Phase timeouts are deterministic and leave no owned child/database process behind.
- Normal installed production lanes finish with meaningful headroom under their deadlines on supported runners.
- Runtime reductions do not remove VCS/sdist, base/PG, restart, historical-artifact, or installed-origin coverage.
- Performance changes preserve package isolation and exact artifact identity checks.
- At least one negative control proves a deliberately stalled phase fails with actionable diagnostics.
Problem
The installed reporting production gates run close to their hard deadlines and provide little or no progress output while the isolated child process is active. This makes a production deadlock, a slow phase, and a runner-capacity problem look alike until the outer timeout expires.
The rc.4 adoption incident in #1195 is the immediate example:
vcs-pg-installed-productionproduced no stdout/stderr for 1,800 secondsThis issue is for test performance and observability. It must not replace or conceal the production concurrency investigation in #1195.
Desired changes
Acceptance criteria