fix(bench): correct matrix-harness reuse measurement, thread guard, resume

[wesleyscholl]

Why

The ollama-vs-squish benchmark matrix (benchmarks/ollama_vs_squish/matrix/) had four issues that blocked a real run. All four are harness-only — no change to squish/ runtime.

Fixes

• Reuse mismeasured. This Ollama build reports the full prompt_eval_count even when the KV prefix is reused, and Squish's reuse counters miss the prefix-slot path. Both engines now fall back to the prefill-time collapse 1 - warm/cold against a cold-prefill reference measured per (system, ctx), keeping the head-to-head apples-to-apples. A sentinel run-index (10_000_000) ensures the cold reference shares no cache prefix with the primed block. (cache_probe.py, cell.py)
• Thread crash. RSSSampler/TemperatureSampler shadowed a threading.Thread internal with self._stop; renamed to _stop_event. (memory.py, thermal.py)
• Governor false-positive. "compressed"/"swap" matched benign "free_swap=0 B" log lines and falsely flagged degradation; narrowed to "swapping"/"paged out". (memory.py)
• Squish metrics URL was /metrics; it is served at /v1/metrics. (systems.py)

Also

• --resume <dir> for run_matrix: cells whose <cell_id>.json already exists are loaded and skipped → crash-safe / overnight-resumable.
• Two unit tests for the Ollama prefill-ratio fallback (count-unreliable → ratio wins; genuine partial count → count wins).

Test

CI=1 .venv/bin/python -m pytest tests/test_benchmark_matrix.py -q   # 58 passed
ruff check / ruff format --check                                    # clean

🤖 Generated with Claude Code