macOS: add a bounded native stack profile and cell-scaling benchmark

[tcconnally]

Problem

Activity Monitor labels the application Python, but native libmpv/FFmpeg and Qt/OpenGL work are charged to that process. Task-level powermetrics proves CPU pressure but cannot identify the hot stack.

Scope

Add a documented, bounded profile and parser that records the exact commit, PID, cell count, decoder per cell, sample stacks, powermetrics task CPU, thermal pressure, GPU residency, frame-pump counters, render timing, and loop lag. Run short churn-disabled profiles at 1, 2, 4, 6, and 8 cells.

Acceptance criteria

• Parser tests use sanitized fixtures and do not require macOS.
• Reports distinguish Python dispatch, native libmpv/FFmpeg, render API, WindowServer, and thermal evidence when the source provides it.
• Missing or permission-denied powermetrics is marked incomplete, never passed.
• The report identifies the first cell count that violates the responsiveness budget.

[tcconnally]

Implemented in PR #81, merged as 300e14c. The native profile parser and matrix CLI now accept sanitized captures or analyze_run() reports and identify the first failing 4/6/8-cell mode. A physical M5 capture is still required to populate the matrix.

[tcconnally]

Native acceptance update

Current main is 4475e0ff8f9207f28f92f0067d9d0ba8b739094d. The parser and bounded profile CLI are present in merged PR #81, but no current M5 host capture is available. sample and privileged powermetrics evidence must be collected on the target Mac; missing or permission-denied native evidence is incomplete, not pass.

The required fixed-corpus 1/2/4/6/8-cell matrix and first failing cell count have not been produced at the current runtime commit. This issue remains open.

[tcconnally]

Post-merge status: main is now b94be1ce3d86f06841ff458e3708d66efacd4ef1 after PR #84. PR #84 changed only the #79 capacity-evaluator contract; it produced no Mac/native measurements and does not change the open native blockers for this issue. The issue remains open.

[tcconnally]

PR #85 merged as 74a4118c6c194c4f07a666c6ffb3112218137829. It fixes the reproduced macOS sample parser defect: current rows such as 2930 Thread_1168979 ... now count as thread records while existing formats remain supported. The captured 8-cell pilot still does not qualify native acceptance: sample is now parser-compatible, but the run used the full 856-item library, lasted about 120 seconds, had no lid-open power samples, p95 loop lag 81.6 ms, one active stall over 100 ms, and paint gaps up to 1970 ms. Issue remains open for fixed-corpus 1/2/4/6/8-cell evidence.

[tcconnally]

Native profile update on main 74a4118c6c194c4f07a666c6ffb3112218137829: the parser now accepts the captured macOS sample format. The existing capture reparses complete with 8 thread records and 120 powermetrics samples; thermal samples were Nominal. The run is still only an 8-cell, roughly 120-second full-library pilot, not the required fixed-corpus 1/2/4/6/8 matrix. Issue remains open.

[tcconnally]

Native pilot update for run 20260905T165329Z/profile-8-docked-clamshell-fixed on commit 2ce6de1c66ce7f4217249c59fa710268c47bc2bc. sample and powermetrics parsing completed with no missing evidence: 120 powermetrics samples, 234 sampled threads, thermal levels Heavy=4/Moderate=32/Nominal=84, and stacks containing libmpv, libavcodec, QOpenGLWidget::paintGL, WindowServer, and mpv_render_context_render. The run used 8 cells and the full 801-item library, not the required fixed 1/2/4/6/8 matrix. Responsiveness also exceeded policy at p95 loop lag 114.5 ms and maximum paint gap 1947.874084 ms. Issue remains open.

[tcconnally]

Native profiling acceptance complete on commit 074258ddb6b1b2e3e1e8bd5cbe09742310c5d99b.

The fixed corpus was the 67-item mv Favorites manifest (0c1b04d2739bd0a151768b2f781644b505e440ad7a4e44562b52ce367895be5e). Churn was disabled. Short phases were completed at 1, 2, 4, 6, and 8 cells with complete sample and privileged powermetrics evidence. All native parsers exited 0 with no missing evidence.

The profiles distinguish Python dispatch, libmpv/libavcodec, QOpenGLWidget::paintGL, mpv_render_context_render, WindowServer, thermal, process CPU, and GPU residency evidence where present. The capacity selector identified 4 cells as the first failing supported mode: 4 failed render-gap; 6 failed p95 loop-lag/render-gap; 8 failed p95 loop-lag/render-gap/loop-stall/freeze gates. The native evidence contract and first-failure reporting criteria are satisfied. This closes the profiling/parser issue; it does not promote any capacity mode or close the separate render/reliability issues.