fix(beat): the GaussianNB "4.9x" was a dev-box number - re-pin to the host that actually runs it

[noahgift]

PMAT-GNB-SPEED-DEFENSE-001. The beat pinned "Measured ~0.20 (apr ~4.9× faster on a 16-core box)" against a 0.50 ceiling — which reads as 2.5× of headroom. The real headroom on the machine that runs this beat is about 20%, and the gate has been tripping: 07-03 0.563, 07-04 0.520, 07-28 0.601.

Measured this session

host	apr	sklearn	ratio
lambda-vector (modern, 48 threads)	~12.0 ms	~56 ms	0.188 / 0.237 / 0.220 (4.2–5.3×)
mac-server — canonical nightly host, IDLE	28.0 ms	70.3 ms	0.398 (2.51×)
mac-server under nightly load (run logs)	—	—	0.278 … 0.601 (straddles 0.50)

This is not decay. git log --since=2026-07-01 -S"GaussianNB" -- crates/aprender-core/src/ is empty; the last impl change was #2181 (2026-06-22) and the ratio has bounced across that entire window. It's a different machine, plus contention.

Why the host moves it so much

apr's GaussianNB is a scalar Rust loop; sklearn's is vectorised numpy. mac-server is a Xeon W-3245 with AVX-512 (avx512f/bw/cd/dq/vl), so numpy closes much of the gap there: lambda → mac-server costs apr 2.3× (12.0 → 28.0 ms) but sklearn only 1.25× (56 → 70.3 ms). On top of that the box runs 17 registered runners on 32 threads, which pushes a busy night the rest of the way over 0.50.

What this does — and does not do

It does NOT weaken the gate. RATIO_CEILING stays 0.50. The idle-host reading (0.398) clears it, so a run over 0.50 remains real signal rather than expected noise. Moving the threshold to make red go away would concede a win we still hold.

It fixes the false baseline, recording each measurement with the host it came from, so nobody re-tightens the gate on a number that machine cannot produce.

It adds a fit/predict phase split, printed every run and included in the failure message — the missing diagnostic. Contention drags both phases up together; an algorithmic regression lands in one. With only a fused total, five red nights told you nothing actionable.

BEAT-SKLEARN-GAUSSIANNB-SPEED: apr=11.675ms sklearn=51.795ms ratio=0.225 (apr 4.44x faster)
BEAT-SKLEARN-GAUSSIANNB-SPEED-PHASES: fit apr=4.666ms skl=15.508ms ratio=0.301
                                    | predict apr=7.078ms skl=37.199ms ratio=0.190

predict is where apr's edge is largest (5.3×) and fit is the weaker half (3.3×) — which also says where the durable fix belongs: SIMD the predict hot loop via trueno to restore real margin on AVX-512 hosts. That remains open under PMAT-GNB-SPEED-DEFENSE-001, deliberately not attempted here because it needs its own before/after on mac-server rather than a dev box.

Note: this beat has been unobservable since 2026-07-26 — the nightly lane died at step 1 on a poisoned uv env cache and fail-fast skipped every later beat (#2326). The 0.601 reading came from re-running the lane after repairing the host.

🤖 Generated with Claude Code