A16: per-detector allocs/event bench + ratchet gate

[trilamsr]

Roadmap

Item A16 from memory/project_horizon_roadmap.md. Stopped mid-session; agent had measured baseline + drafted gate before stop.

Scope

Per-detector allocs/event bench + regression gate via make bench-check. Baselines measured this session (Apple M1 Max):

Detector	N	allocs/op	allocs/event
pod_evicted	1024	15635	15.27
nccl_hang	800	3177	3.97
xid_correlation	200	819	4.09
hbm_ecc	200	1622	8.11
thermal_throttle	240	892	3.72
pcie_aer	200	2722	13.61

All exceed the 2 allocs/event NORTHSTAR target. Plan: ratchet ceilings to measured baseline + file per-detector optimization issues. Ceiling ratchets DOWNWARD as optimization PRs land; never up without adversarial review.

Deliverables

• module/pkg/patterns/{detector}_bench_test.go per detector (5 files; pod_evicted exists already)
• scripts/bench-registry.sh: extend bench_entries + add allocs_gate array (bench_name|N|ceiling)
• scripts/bench-allocs-check.sh (new) — reads allocs_gate, extracts median allocs/op from module/pkg/patterns/testdata/bench-baseline.txt, fails on exceed
• Wire into make bench-check
• Regenerate module/pkg/patterns/testdata/bench-baseline.txt with count=10 for benchstat
• Per-detector optimization tracking issues (5 — pod_evicted skip; already at native limit pending re-investigation)

Initial ceilings (measured baseline + 1 slack)

• pod_evicted: 16 / 1024 = 16
• nccl_hang: 4 / 800 = 4
• xid_correlation: 5 / 200 = 5
• hbm_ecc: 9 / 200 = 9
• thermal_throttle: 4 / 240 = 4
• pcie_aer: 14 / 200 = 14

Acceptance

• make bench-check green w/ committed baseline
• Synthetic 10x slowdown fixture trips the gate
• Per-detector optimization issues filed before ceiling-ratchet PRs

Adopt-over-build

• stdlib testing.B + -benchmem only

Out of scope

• Detector hot-path optimization itself (separate PRs per optimization issue)