refactor(patterns): introduce Detector registry seam

[trilamsr]

Summary

Wave-end audit flagged the patterndetectorprocessor fanout site as an unmet refactor: ConsumeLogs hand-rolled dispatch for every shipped detector (12 today: 7 inline + 5 wrapped), so adding pattern #13 required editing the fanout body — not registering a new entry. Past the rule-of-three by 9x.

This PR introduces a minimal Detector registry seam:

• module/pkg/patterns/detector.go: new Detector interface (PatternID() string) + Registered slice that pins all 12 detector pointers. Each *Detector struct gets a one-line PatternID() method.
• module/pkg/patterns/detector_test.go: TestRegistered_PinsAllPatterns (exact PatternID set + count), TestRegistered_UniquePatternIDs, TestRegistered_NonEmptyPatternIDs. Drift gate — accidental drops fail in CI.
• patterndetector.go: introduces detectorRunners []detectorRunner closure list iterated by ConsumeLogs. ConsumeLogs body drops from ~77 lines to 12. Adding pattern [general] M1.6 hardening: lint + coverage + panic recovery + pipelinebuilder #13 = one append to Registered + one append to detectorRunners, no fanout-site edit.

Design decision: metadata-only interface

The Detector interface is intentionally PatternID() string only — not a uniform Evaluate method. Each detector's Evaluate signature is intrinsically heterogeneous (different input record shapes — events+nodeConds, ncclRecs, xidRecs+events, etc. — and different verdict types). A uniform Evaluate would force a lossy any-typed contract that the typed test suite has been fighting for 12 patterns. The closure-per-detector approach keeps the typed Evaluate calls at their concrete-runner sites while letting the registry pin identity + iteration.

Behavior preservation

• Same telemetry vocabulary: PodEvicted and IBLinkFlap still IncVerdict with string(v.Confidence) (they gate on partial); the other 5 inline detectors still pass "". The 5 wrapped runners still tick inside their own helpers (unchanged).
• Same emission order: detectorRunners is declared in the legacy emission order.
• Same partial-confidence gating: emitPodEvicted and emitIBLinkFlap preserve the !emitPartial skip.

Test plan

• cd module && go build ./... clean
• cd module && go test ./pkg/patterns/ green (incl. 3 new pin tests)
• cd module && go test ./processor/patterndetectorprocessor/ green except pre-existing test: synthetic-2026-06-multi-rank-disk-pressure fixture mis-labelled as negative #497 (TestPatternDetector_NegativeFixturesEmitNoVerdicts/synthetic-2026-06-multi-rank-disk-pressure, fixed in Lane J)
• make lint clean (0 issues)
• make vet, go mod verify, attribute-namespace-check all green (pre-push hook)

LoC delta

• +321 / -79 across 3 files.
• ConsumeLogs body: 77 → 12 lines.
• Growth is in: registry plumbing (164 lines, mostly comments + the pin tests), runner closures (one per detector). The seam earns its bytes — adding pattern #N is now O(append) instead of O(edit-fanout).

Closes-the-loop

Closes wave-end-audit next-wave item #2 (pattern registry seam).

- refactor(patterns): introduce `patterns.Detector` interface + `patterns.Registered` slice. The patterndetectorprocessor now iterates a registry-driven runner list instead of hand-rolled fanout — adding a new pattern is one append, not a processor edit. Behavior-preserving; no operator-facing change.