perf: complete macOS render, decoder, and capacity controls

[tcconnally]

Summary

This is the coordinated follow-up for the macOS performance roadmap:

• Add an explicit macOS-only HYPERWALL_RENDER_PROFILE=low-cost tier without changing Windows/Linux defaults.
• Export per-resource decoder attempts, activations, software fallbacks, exhausted recovery, and quarantine.
• Parse p95 loop lag and active stalls at or above 100 ms.
• Add a fail-closed 4/6/8-cell capacity evaluator with duration, responsiveness, render-gap, CPU, decoder, audio, A/V, and transport evidence.
• Make the profile CLI accept normalized profiles or analyze_run() reports and select the highest passing mode.
• Update operator documentation and the roadmap to describe the usable workflow.

Operational intent

The goal is to make the wall behave predictably under pressure: reduce avoidable GUI queue work, keep one bad resource from changing wall-wide decoder policy, and avoid silently promoting an unsupported cell count. The default HQ render path remains unchanged until a fixed-corpus comparison proves the cheaper tier is acceptable.

Validation

• python3 tests/run_all.py: all suites green.
• Focused platform, decoder, diagnostics, profile, and capacity tests pass.
• Python compilation and git diff --check pass.

Hardware boundary

This PR provides the controls and reporting needed for the M5 run. It does not claim that 8-cell playback is production-ready, does not make VideoToolbox a global default, and does not use the rejected 30 fps / 25 Mbps live-transcode shaping. The physical 4/6/8-cell M5 matrix and fixed-corpus visual comparison remain qualification work.

Tracking

Refs #75
Refs #76
Refs #77
Refs #78
Refs #79