New Skill: hrr-replay-analysis

[andraghetti]

Summary

Adds hrr-replay-analysis, the skill the catalog has advertised as _planned_ since #94. It replays a HIP Record and Replay (HRR) archive on the host GPU and reports what went wrong as a structured finding: fault class, fault address, failing event index, and the implicated kernel.

An HRR archive is a deterministic record of a real HIP workload, so a customer failure can be reproduced from the recording alone, without their application, source or data. The skill stops at the finding and hands off; kernel patching, rocgdb and fix verification are deliberately out of scope.

Layout

skills/hrr-replay-analysis/
├── SKILL.md
├── reference.md
├── examples.md
├── skill-card.md
├── scripts/run_hrr_replay.sh
├── scripts/analyze_replay_finding.py
└── evals/
    ├── evals.py
    └── fixtures/

Testing

Structural gates:

• ./.github/scripts/check.sh: 7 skills, 0 errors, Cursor and Codex manifests in sync
• Offline link check over the skill's markdown, 0 errors
• SkillSpector scan: no HIGH or CRITICAL findings (2 MEDIUM for subprocess, both advisory)

End to end on a gfx950 host, against three real archives:

• Clean workload. An offline vLLM inference capture (185,469 events, 13,233 kernels, 33 code objects, 2 threads) replays clean in 39 s with all 32 device-to-host checks exact, and is reported as a pass.
• Crashing workload. A capture whose workload faults reproduces the memory access fault on replay, and the finding carries the fault class, the fault address and the faulting kernel.
• Unreadable archive. An archive whose format version the playback build cannot read is reported as unreadable, not as a workload failure.

That validation changed the skill. The clean workload was reporting a hipBLASLt GEMM as its kernel, matched out of the archive's own kernel listing, which next to a pass reads as a culprit; a clean replay now implicates nothing. The crashing workload was reporting no kernel at all, because a memory fault tears the process down before the failing dispatch is attributed, so the archive's kernel table is now used as an explicitly-marked fallback when it holds exactly one kernel. Both cases are covered by fixtures. The full list of parser and device-masking fixes is in the commit message.

Behavioral evals come in two tiers. Four tests run against replay logs checked into evals/fixtures/, so they grade the agent's reasoning (right fault class, right address, right kernel) on a runner with no GPU, and they include an unreadable-archive guard and a false-activation screen. A fifth test replays a real archive end to end and skips unless /dev/kfd, an hrr-playback binary and HRR_EVAL_ARCHIVE are all present, so it stays opt-in rather than failing where no GPU is available.