macOS: define M5 cell-capacity policy and promotion gate

[tcconnally]

Problem

An 8-cell heterogeneous direct-play wall on a fanless 16 GB M5 has not passed a clean long-run media gate in either software or hardware-decode mode. The product needs an explicit capacity policy instead of repeated decoder toggles.

Scope

• Define 4-, 6-, and 8-cell operating modes.
• Promote the highest measured passing mode and fail closed when evidence is missing.
• Keep 8 cells experimental until the complete 30-minute fixed-corpus gate passes.
• Document that the rejected runtime 30/25 live-transcode shaping is not an option; use fewer cells or a pre-normalized corpus.

Acceptance criteria

• Each promoted mode has CPU, loop-lag, render-gap, freeze, decoder, audio, and duration-coverage evidence.
• README and docs/performance-roadmap.md separate harness health from media health.
• No claim of 8-cell readiness is made without the exact gate evidence.

[tcconnally]

Implemented in PR #81, merged as 300e14c. Capacity selection is now fail-closed across 4/6/8 cells with coverage, p95 loop lag, render gap, CPU, decoder, audio, A/V, and transport evidence. The physical M5 matrix still determines which mode can be promoted.

[tcconnally]

Native acceptance update

Current main is 4475e0ff8f9207f28f92f0067d9d0ba8b739094d. No current 4/6/8-cell 30-minute fixed-corpus matrix exists with duration, responsiveness, media, and independent AC/lid/sleep power evidence. The previous 8-cell artifacts are blocked by current analysis and do not establish a promoted mode.

The capacity evaluator remains fail-closed. Until the target Mac run is available, promoted cells: none; 8 cells remains experimental. This issue remains open.

[tcconnally]

PR #84 merged the capacity-evaluator power-evidence contract at b94be1ce3d86f06841ff458e3708d66efacd4ef1. It did not complete the required native M5 4/6/8-cell matrix, so this issue is reopened and remains open. Promoted cells: none; 8 cells remains experimental pending current fixed-corpus duration, media, and independent AC/lid/sleep evidence.

[tcconnally]

Independent exact-tree review of PR #84 passed at candidate head 00162bdbff33c66ffb639118c5a497761f8f3d87 with diff fingerprint 5dfccf77128002a191ed9bab6b96f2ec5ca46b858ea05ec76ca8b51e0991362c. The review verified that missing, WARNING, and BLOCK power gates fail closed, and that only an exact PASS gate can promote a normalized profile. It noted one LOW coverage item: missing and WARNING do not yet have separate named test functions, although the boundary probe and full suite covered the behavior. Native Mac qualification is still unavailable, so this issue remains open.

[tcconnally]

Capacity update on main 74a4118c6c194c4f07a666c6ffb3112218137829: the 8-cell pilot had duration coverage for only about 120 seconds and the independent power gate is BLOCK because AC was observed but lid-open samples were zero. No mode is promoted; 8 cells remains experimental. A fixed-corpus, lid-open/awake 4/6/8 matrix is still required. Issue remains open.

[tcconnally]

PR #86 merged as 56500a90d2ebc2a45fb57f5aa8ac1c891814c61c. Power evidence now accepts the supported docked-clamshell mode: AC power and sleep assertions plus verified external-display evidence with a closed lid. Open-lid evidence remains valid, and missing external-display evidence still fails closed. The prior 8-cell run remains non-qualifying because it predates this marker, used the full library, was only about 120 seconds, and failed responsiveness gates. No cell mode is promoted; issue remains open.

[tcconnally]

Native pilot update: 20260905T114942Z/profile-8-docked-clamshell completed on main 56500a90d2ebc2a45fb57f5aa8ac1c891814c61c, but wrote external_display_observed=0 even though system_profiler SPDisplaysDataType listed two LG ULTRAGEAR+ displays under Displays:. The macOS output omitted both Display Type and Connection Type, so the docked-clamshell power gate correctly remained blocked. The run is exploratory only: 8 cells, full 856-item library, about 120 seconds, and no capacity promotion. PR #87 merged as 2ce6de1c66ce7f4217249c59fa710268c47bc2bc; its parser now recognizes named displays in this macOS output while retaining legacy markers. A fresh run on the merged commit is required.

[tcconnally]

Native pilot update for run 20260905T165329Z/profile-8-docked-clamshell-fixed on commit 2ce6de1c66ce7f4217249c59fa710268c47bc2bc. The docked-clamshell power gate now passed with external_display_observed=true, docked_clamshell_observed=true, 25 power/assertion samples, and 25 lid-evidence samples. The 8-cell pilot still used the full 801-item library and lasted about 119.812 seconds; p95 loop lag was 114.5 ms, there were 126 stalls at least 100 ms, and maximum paint gap was 1947.874084 ms. No cell count is promoted; 8 cells remains experimental pending fixed-corpus duration and responsiveness gates. Issue remains open.

[tcconnally]

Fixed-corpus capacity matrix update on commit 074258ddb6b1b2e3e1e8bd5cbe09742310c5d99b.

Corpus: Emby mv Favorites, 67 items, manifest SHA-256 0c1b04d2739bd0a151768b2f781644b505e440ad7a4e44562b52ce367895be5e. All five short phases used HYPERWALL_SOAK_FILTER=favorites before cell startup and had complete docked-clamshell power evidence plus complete native sample/powermetrics captures.

Run IDs:

• 20260905T193640Z/profile-1-favorites-initial
• 20260905T195016Z/profile-2-favorites-initial
• 20260905T195747Z/profile-4-favorites-initial
• 20260905T200732Z/profile-6-favorites-initial
• 20260905T203318Z/profile-8-favorites-initial

The capacity selector returned BLOCK, selected cells null, and first failing supported mode 4. Failures: 4 cells exceeded the render-gap gate; 6 cells exceeded p95 loop-lag and render-gap gates; 8 cells exceeded p95 loop-lag, render-gap, loop-stall, and freeze gates. No mode is promoted. Eight cells remains experimental. Issue remains open.