fix: reject suspended soaks and restore M5 cache

[tcconnally]

Summary

• restore the small-mac cache profile to 256 MiB/cell and 2 GiB aggregate while retaining the software-decoder default
• run macOS soak phases under caffeinate -dims to prevent idle sleep from invalidating the measurement
• add a fail-closed duration-coverage gate: active soak duration must cover at least 95% of the requested interval

One-hour screen evidence

The 128 MiB/cell run was a clear BLOCK:

• 53 cache-starvation freezes / 426.8 seconds
• 34 playback errors, 3 video decode errors, 83 A/V desync warnings
• 320 loop stalls; 23 exceeded 500 ms; max 2640 ms
• hardware decode failures remained zero, validating the software-decoder half of perf: tune 16 GiB macOS playback profile from overnight soak #71

The process ran for one wall-clock hour, but the manifest recorded only 2673.078 seconds (74.25%). hyperwall.log, vm_stat, and nettop all stopped together for about 14m45s, consistent with host suspend. The old analyzer accepted this incomplete interval; the new gate blocks it.

Greg-side correlation for the run showed only successful HTTP status classes and 11 completed transcodes with no conversion failures. This does not support relocating Emby as the first remediation.

Verification

• RED observed for 256 MiB cache expectations, anti-sleep launcher contract, and duration gate/wiring
• python3 tests/run_all.py — OVERALL PASS
• bash -n soak_wall.sh — PASS
• python3 -m compileall -q hyperwall tests scripts — PASS
• git diff --check — PASS
• replaying the attached run through the new analyzer produces duration_coverage=BLOCK with 0.7425 coverage

Next native gate

Repeat the 60-minute, 8-cell software-decoder screen with default cache settings. Do not promote to overnight until the full active duration is captured and media gates materially improve.