Integrate capability-aware Apple Silicon runtime

[Paul (nuclearspike)]

Replaces #138, which I closed to correct the contents of the branch. Same technical change, clean history.

Summary

• Consolidates draft PRs Fix silent loss of completed Extend results #132, Clarify Pro model availability in local mode #133, Fix timeline gap frame conditioning #134, and Release reused diffusion cache before VAE decode #136 without modifying their source branches.
• Adds capability-aware Fast T2V/I2V routing: auto uses MLX only for local text encoding, cached BF16 weights, and the exact clean 0.14.20.dev1 runtime at revision 3171bac4ba901c0237faea2678c34034b37abc2a. Prepared embeddings, A2V, Retake, Extend, IC-LoRA, and image generation stay on Torch.
• Runs MLX in a fresh process group for every job. The backend never imports MLX; cancellation kills the whole group and process exit is the allocator teardown boundary.
• Resolves local Fast dimensions before conditioning or inference. Distilled 540p uses the budget-bounded 896x512 two-stage grid (512x896 portrait), removes hidden per-axis rounding, reports the resolved pair in the completion API, rejects off-grid MLX commands, and makes strict QA require API dimensions to equal ffprobe output.
• Persists a prompt-safe JSONL profile for every MLX job and exposes live/last-job allocator active, cache, peak, phase, terminal status, path, timestamp, and exact runtime identity through the existing telemetry API and Settings UI. Incomplete tails, invalid timestamps, non-finite values, and killed-child cancellation are handled without breaking telemetry.
• Adds an fcntl-based shared Metal lease so a second local GPU consumer on the same machine cannot run concurrent Metal work, plus process/system/MPS/MLX memory telemetry, capability routing and exact provenance in Settings, and a strict HD performance evidence harness that captures the post-child terminal profile.
• Keeps q8 expert-only with a visible quality warning; auto never selects it. Low-RAM MLX enables runtime auto-tiling.

Evidence

• Backend: 596 passed.
• TypeScript and Pyright: clean.
• OpenAPI generated bindings and drift check: clean.
• Production frontend and Electron bundles: passed.
• Full macOS production build from the integration base: passed; signed app, DMG, and ZIP produced.
• Live Apple M5 Max policy smoke: exact MLX runtime verified, BF16 cache detected (139,683,517,037 bytes), auto selected MLX for eligible Fast local-text work, backend idle RSS 497 MiB.
• Pre-fix production smoke proved the former mismatch: resolver 960x544 versus actual sidecar/output 960x512. The final post-fix profiled smoke is delegated to QA.
• PR Release reused diffusion cache before VAE decode #136 MPS 1280x704x121 benchmark: median 220.570s to 209.465s (-5.04%); peak 44,504 MiB to 41,713.5 MiB (-6.27%); identical output.
• Persistent MLX worker was rejected: retained about 50.8 GiB and request 2 was slower.

Benchmark evidence under backend/performance_runner/artifacts/ has machine-specific paths and hostnames redacted.