Address reviewer A- trim items on #456:
1. displayPodName: drop optimization-impact comment (function inlined
   in pod_evicted; kept exported for xid_correlation use).
2. Evaluate: collapse 9-bullet allocation-changelog into one-line
   NORTHSTAR pointer + per-file note.
3. Consolidate fmt->strconv narrative into one top-of-Evaluate note;
   drop the per-site repetition across five render helpers.
4. containsFold: replace 'equivalent to' note with explicit ASCII-only
   safety guard (anchors + kubelet eviction messages are English).

No behavior change: allocs/op = 1943 unchanged, golden replay
byte-identical, bench-allocs-check still PASS.

Verified finding 5: scripts/bench-registry.sh xid_correlation
ceiling already 2 (post-#448), no change required.

Signed-off-by: Tri Lam <tree@lumalabs.ai>

…434) (#456)

## Summary

Closes #434. Profile-driven optimization of
`PodEvictedDetector.Evaluate` drops per-event allocations from **15.27 →
1.90**, clearing the **NORTHSTAR (2)** floor. Bench ceiling lowered 16 →
2 in `scripts/bench-registry.sh::allocs_gate`;
`bench/detectors/baselines.json` updated. Same recipe as #438
(nccl_hang) and #448 (xid_correlation).

## Before / after

`go test -bench=BenchmarkPodEvictedDetector -benchmem -count=10
-benchtime=500ms ./bench/detectors/` (Apple M1 Max):

| metric | before | after | delta |
|---|---|---|---|
| allocs/op | 15635 | 1943 | **-87.6%** |
| B/op | 970020 | 473400 | -51.2% |
| allocs/event | 15.27 | **1.90** | hit NORTHSTAR |

Stability: 10 consecutive bench runs all measured 1943 allocs/op (CV =
0%).

## Root cause (named)

Per-evicted-pod the previous detector allocated ~19 objects. Profile
(`-memprofile`, `-alloc_objects`) attributed them:

1. **`fmt.Sprintf` ×5 in `buildVerdict` + helpers** — 60% of pre-fix
allocs. Sprintf boxes each arg into an `any` slice + scans the format
string into a fresh internal buffer; both escape. Replaced with
`strconv.AppendInt` + manual byte builders.
2. **`strings.ToLower(note)` in `pressureFromNote`** — 1 alloc per
evicted pod (the lowercased copy). Replaced with a zero-alloc
`containsFold` ASCII-fold scan (all anchors are already lowercase).
3. **`time.Format` in `formatTimestamp`** — Sprintf-style alloc per
headline. Replaced with `t.AppendFormat(buf, time.RFC3339)` writing into
the shared scratch buffer.
4. **Unsized `condIdx` map + per-bucket `append` growth** —
`make(map[K]V)` defaults to 8 buckets; `append(nil, r)` grows each
bucket. Pre-sized to `len(recs)`.
5. **Per-verdict `make([]EvidenceRef, 1, 2)`** — 1 alloc per verdict for
the trail slice header. Replaced with ONE contiguous `trailBacking :=
make([]EvidenceRef, 2*len(events))` sliced cap=2 per verdict.
6. **Escaping `make([]byte, dynamic-cap)` per builder call** — escape
analysis can't prove the dynamic capacity fits on stack, so each builder
allocated TWO objects (buf + the `string(buf)` cast). Replaced with a
per-Evaluate `scratch *[]byte` reused across every builder (each resets
to `(*scratch)[:0]` and appends), collapsing to ONE alloc per call (the
irreducible string cast).
7. **Redundant per-condition re-render** — 820 evictions joining onto 64
conditions re-rendered the same UID + description + remediation strings
on every join (2460 redundant allocs). Pre-rendered once at index time
into a new `indexedNodeCond` struct.
8. **Partial-path remediation re-render** — partial-path evictions (no
joined condition) re-rendered the `"On node X: <prose>"` remediation per
pod. Added a `(node, pressure) → string` cache so burst evictions on the
same node share one alloc.
9. **`sort.SliceStable` reflection** — both `indexNodeConds`'s
per-bucket sort and the outer verdicts sort used reflection-based
`sort.SliceStable` (allocates a swapper per call). Replaced with
`slices.SortStableFunc` (generics, no reflection); buckets of length ≤1
skip the sort call entirely.

## Top profile entries (memprofile -alloc_objects)

**Before:**
```
60% buildVerdict (5 fmt.Sprintf calls per eviction)
22% fmt.Sprintf
 6% nodeConditionDescription
 4% annotateRemediationWithNode
 3% time.Time.Format
 4% strings.ToLower
```

**After:**
```
42% podEventDescription   (1 alloc/call — irreducible string buf cast)
36% renderHeadline        (1 alloc/call — irreducible string buf cast)
 7% annotateRemediationWithNode  (partial-path cache misses only)
 5% nodeConditionDescription     (index-time pre-render, 64 calls total)
 3% nodeConditionUID             (index-time pre-render, 64 calls total)
```

The two per-verdict string allocations (headline + pod_event
description) are now the irreducible floor: each is a distinct verdict
field that downstream `json.Marshal` requires as separate strings.

## Verdict semantics — unchanged

Byte-identical headlines, remediation, UIDs, descriptions, JSON
envelope. Confirmed via:

- `module/pkg/replay/pod_evicted/canonical/golden.json` round-trips
unchanged.
- All `TestPodEvicted*` unit tests pass (negative fixtures,
deterministic order, partial path, out-of-window, future transition
excluded, empty node message, remediation pins node name, schema
conformance + drift rejection).
- `TestPodEvictedVerdict_SchemaConformance` (canonical full + partial)
passes.

## Gate state

| gate | result |
|---|---|
| `scripts/bench-allocs-check.sh` (hard absolute ceiling) | PASS at new
ceiling=2 |
| `scripts/bench-check-detectors.sh` (10% delta soft gate) | PASS (no
detector regressed) |
| `go test -race ./module/pkg/patterns/ ./module/pkg/replay/
./module/processor/...` | PASS |
| `go vet ./module/pkg/patterns/` | clean |
| pre-commit (golangci-lint, vet, mod verify, attribute-namespace-check)
| PASS |

```release-notes
pod_evicted detector now allocates 1.90 allocations per event (down from 15.27), hitting the v0.3.0 NORTHSTAR floor of <=2 allocs/event. Joins #418 (nccl_hang) and #417 (xid_correlation) — three of the six per-detector tracking issues at or below NORTHSTAR.
```

Closes #434
Refs #302

## Test plan

- [x] `go test -race -count=1 ./module/pkg/patterns/
./module/pkg/replay/ ./module/processor/...` PASS
- [x] `go vet ./module/pkg/patterns/` clean
- [x] `bash scripts/bench-allocs-check.sh` PASS (pod_evicted 1.90/ev <=
ceiling 2; all 6 detectors at-or-below ceiling)
- [x] `bash scripts/bench-check-detectors.sh` PASS (no detector
regressed >10%)
- [x] `go test -bench=BenchmarkPodEvictedDetector -count=10
-benchtime=500ms` stable at 1943 allocs/op (CV=0%)
- [x] Golden replay
(`module/pkg/replay/pod_evicted/canonical/golden.json`) round-trips
unchanged
- [x] Rebased onto origin/main (resolved scripts/bench-registry.sh
conflict with #448's xid_correlation ratchet — both ratchets shipped
together in this branch's view)

---------

Signed-off-by: Tri Lam <tree@lumalabs.ai>

## Summary

`scripts/bench-registry.sh::bench_entries` covered only
`BenchmarkPodEvictedDetector` and the pyspy parser benches, so the **%
delta regression gate** (`scripts/bench-check.sh` via
`bench-check-all.sh`) missed five of the six detectors. The absolute
allocs/event ceiling (`allocs_gate`) caught a detector that drifted past
its hard ceiling, but a detector that drifted +50% while staying under
the ceiling would ship silently. This PR closes the gap by registering
all six detectors against both gates.

## Root cause

The registry was authored when `pod_evicted` was the only detector with
a bench. The five later detectors (added in
#417/#418/#434/#438/#448/#456) were plumbed into `allocs_gate` (absolute
ceiling) but the parallel % delta gate was never extended — a historical
artifact, not an architectural defect. Fix: extend `bench_entries` to
cover the missing five.

## What changed

- `scripts/bench-registry.sh`: added one entry covering five detectors
against `./bench/detectors/`, single-regex / single `go test` invocation
(~5s incremental vs ~30s if split into 5 entries).
- `bench/detectors/testdata/bench-baseline.txt`: generated baseline
(count=10 × 500ms, Apple M1 Max). Hardware-invariant signals (B/op +
allocs/op) pin the gate; sec/op stays advisory.
- Existing ceilings in `allocs_gate` unchanged. Existing baselines
unchanged.

## Mutation verification

Bumped `PCIeAERDetector` allocs baseline down by ~18% (524 → 430):

```
PCIeAERDetector-10  430.0 ± 0%  524.0 ± 0%  +21.86% (p=0.000 n=10)
REGRESSION: the following benchmarks exceeded the 10% threshold vs baseline:
  PCIeAERDetector-10  +21.86% (p=0.000 n=10)
```

`scripts/bench-check-all.sh` exit 1. Revert → exit 0. Gate fires on
regression, stays clean on baseline.

## Coverage gap context

Cross-link: #302 (allocs/event rollup), #417 (xid_correlation
NORTHSTAR), #418 (nccl_hang NORTHSTAR), #434 (pod_evicted NORTHSTAR),
#438/#448/#456 (sibling perf PRs). Every detector that hit NORTHSTAR
under the absolute ceiling now also has a relative-regression gate.

## Test plan

- [x] `make bench-check` exit 0 (covers all 6 detectors now)
- [x] `make bench-allocs-check` exit 0
- [x] `make bench-detectors-check` exit 0 (soft-gate, unchanged)
- [x] Mutation: lower baseline → exit 1; revert → exit 0
- [x] Lint + vet (pre-commit hook): 0 issues

```release-notes
ci(bench): % delta regression gate now covers all six detectors (#302). xid_correlation, hbm_ecc, nccl_hang, thermal_throttle, and pcie_aer were previously gated only by the absolute allocs/event ceiling — they now also fail builds on >10% allocs/op or B/op drift vs the committed baseline.
```

Signed-off-by: Tri Lam <tree@lumalabs.ai>