[woksin]

Added

• trial in config.local.json marks a trial window (name, arm, start). Every phase record carries it, and pi-phase report --trial NAME summarises each window and arm: throughput while queued, p50/p95 queue-plus-run time by kind, timeouts, swap growth, peak memory, CPU time and workload mix. (part of pi-phase: measured concurrency, hang-timeout and fast-admission trials #93)
• hang_timeout gains an observe mode and a labels scope. Observe records which phases the hang timeout would apply to without changing their command, and applied phases that end on a hang signature are recorded as hang kills. (part of pi-phase: measured concurrency, hang-timeout and fast-admission trials #93)
• Build and test phases record a fast-lane shadow verdict, the predicted against actual duration with warm or cold worktree, without changing admission. pi-phase report shows its prediction error and how its verdicts compare with today's fast-lane decisions. (part of pi-phase: measured concurrency, hang-timeout and fast-admission trials #93)
• Started phases record the CPU seconds used by their child processes. (part of pi-phase: measured concurrency, hang-timeout and fast-admission trials #93)

Fixed

• A dotnet started by an explicit path, which bypasses the MSBuild worker cap, is now recorded and priced as using all cores instead of the cap. (part of pi-phase: measured concurrency, hang-timeout and fast-admission trials #93)

[woksin]

Reviewer notes: full review round 1 found one blocker (trailing-Z timestamps on Python < 3.11), fixed in fef1600; quick round 2 passed. pi-phase suite: 565 tests passed at 749709a; after the fix the suite hit 3 multi-process timing failures at host load ~160 (queue order, two slots, worktree lock), all passing on targeted rerun; the fix only touches timestamp parsing. Trials (baseline, hang 240s subset, 5x3 vs 4x2) run after install and are tracked on #93.