fix(brain): calibrate Metal from complete runtime allocations

[Draek2077]

What changed

• Budget Metal models against the reported recommended working set with a 0.25 GiB reserve, and allow fits below 4096 context down to 1024.
• Calibrate with two contexts that fit the estimated budget without first loading the saved hosting profile.
• Temporarily request runtime verbosity 5 and capture complete Metal allocations, including the MTL0 memory breakdown that llama.cpp b11181 emits when the sample server closes. Reject incomplete lines rather than saving a partial measurement.
• Show component costs and clearer budget labels in the model profile.

Evidence and limits

• The Mac mini's 0.9.23 logs show MTL0 (Apple M4 Pro) self allocation of 15,714 MiB and a compute buffer of 136.2813 MiB. The existing parser only accepted older Metal ... buffer size = lines, so calibration rejected the load.
• Seven focused test files pass (122 tests). Brain and app typechecks, lint, formatting, and the full pre-commit typecheck pass.
• This candidate has not run on the Mac mini. A live calibration and inference trial is still needed before the next stable release.