fix: make soak evidence self-validating

[tcconnally]

Summary

Apply the lessons from the 2026-08-24 controlled Hyperwall soak:

• Add --expected-cells and a fail-closed analyzer gate so report directory labels cannot misrepresent the actual grid.
• Record aggregate and per-cell cache ceilings in playback policy/plan telemetry.
• Count direct, server-transcode, and prefetch plan decisions across the entire application log instead of relying only on the final cell snapshot.
• Record current resident memory separately from POSIX ru_maxrss peak high-water measurements.
• Document the corrected soak procedure and evidence semantics.

This is observability/evidence hardening only. It does not change the direct-vs-transcode policy.

Verification

• python3 tests/run_all.py — overall PASS.
• Reliability: 73 passed.
• Diagnostics: 67 passed.
• Soak telemetry: 5 passed.
• Playback plan: 11 passed.
• Runtime current-RSS smoke: passed.
• Linux GUI-dependent lanes remain explicitly skipped where PyQt6 is unavailable; target macOS/Windows validation remains required.

Source evidence

The attached soak archive showed all runs labeled 4cell-* actually contained 8 cells. This PR makes that mismatch fail closed on future runs.