metrics: record MLX nax-capability in run identity

[Pushkinist]

Summary

• Stamp the metallib nax-kernel scan crates/rmlx-mlx/build.rs already computes (fix(mlx): gate the loud nax-kernel warning on Neural-Accelerator-class hosts #262's kernels_present) into a new RMLX_MLX_NAX compile-time constant (present/absent/unknown) — not a second detection path, the same scan the missing-kernel build warning already uses.
• Thread it into run identity: rmlx-cli::main() calls rmlx_metrics::identity::set_mlx_nax(rmlx_mlx::NAX_CAPABILITY) once at startup (same process-wide one-shot pattern as install_rotor_qjl/install_planar_fused_qk), because rmlx-metrics cannot depend on rmlx-mlx (whose build script hard-requires a working Homebrew MLX/mlx-c install — a dependency this generic, cross-backend metrics crate must not carry) and cargo:rustc-env only reaches the compiler invocation of the crate whose build script set it, never a different crate's.
• New migration 004_events_mlx_nax.sql adds events.mlx_nax (nullable, append-only); EventRecorder::record now populates it alongside backend_version/build_profile.
• observations / the §8.5 ingest shape is left untouched — that table's identity fields are strictly binary-known-or-caller-supplied-via---git-sha-style flag, and mlx_nax doesn't fit either slot as a forced column there; a bench script that wants it can already carry it in RunRecord.notes (free-form).
• Docs: docs/METRICS_DB.md (new column + why the propagation goes through a runtime setter), docs/FFI.md (points at the new run-identity use of the scan), docs/PERF_BASELINE.md (new top-of-file caveat: prefill numbers recorded ~2026-07-13 through the pin restoration on 2026-07-17 are suspect — 2-3.7x inflated, decode unaffected).

#275's historical-row backfill — declined, not done

#275's title also asks to "annotate post-Jul-13 no-NAX prefill rows in runs.db." This PR does not touch runs.db. events is explicitly append-only with no UPDATE path (migration 003_events_identity.sql's stated contract), and CLAUDE.md lists metrics-table mutation as an Ask-before item. Retroactively writing mlx_nax into rows that never recorded it at insert time would fabricate identity the binary didn't know at the time. Documented instead of mutated — see the caveat sections above. The coordinator should confirm with the maintainer whether any other form of historical annotation (e.g. a documented exclusion range in downstream queries) is wanted; this PR takes no action on the DB itself.

Closes #275 (the run-identity half; the historical-row half is intentionally left to the maintainer per the above).

Test plan

• Unit: EventRecorder::record writes a non-NULL, self-consistent mlx_nax (crates/rmlx-metrics/src/events_tests.rs::record_populates_mlx_nax)
• Unit: migration 004 adds the column, applies cleanly + idempotently after 003 (schema_tests.rs::events_table_has_mlx_nax_column, existing init_schema_idempotent generically covers ordering/idempotency since MIGRATIONS is a plain ordered slice)
• Unit: set_mlx_nax/first-writer-wins semantics (identity_tests.rs::mlx_nax_defaults_to_unknown_then_takes_the_first_set_value)
• Unit: tri-state string mapping (crates/rmlx-mlx/tests/mlx_pin.rs::nax_capability_str_matrix)
• Real run: rmlx baseline --model gemma-4-e2b-it-mxfp8 --kv-quant none --prompt-tokens 4096 --max-ctx 8192 --record against a scratch RMLX_HOME → SELECT DISTINCT mlx_nax FROM events returns present (this dev box's pinned MLX), confirming the real end-to-end plumbing, not just the unit tests. Arch-independent metrics plumbing, so one real model is sufficient (not the ≥2-arch perf-proof rule, which is for perf/kernel claims).
• make ci: fmt/clippy/deny/audit/ci-metrics/gate-scripts all pass; full cargo test --workspace (unit+integration+doc, every crate) ran to completion with 0 failures — the background invocation was killed by the harness mid doc-tests (external timeout, not a test failure) after every unit/integration test in the workspace had already passed, so deny/audit/ci-metrics/the six gate scripts were re-run standalone afterward and are all green.

🤖 Generated with Claude Code