feat(sdc): pattern-13 silent data corruption detector

[trilamsr]

Summary

Ships pattern #13 (silent data corruption) end-to-end per the spec at docs/patterns/13-silent-data-corruption.md — the hardest pattern in the 15-set, where the run completes, loss looks normal, but downstream eval shows degraded model quality.

• Library: SilentDataCorruptionDetector in module/pkg/patterns/silent_data_corruption.go consumes EvalAccuracyRecord + SDCCounterRecord typed projections and emits SilentDataCorruptionVerdict. Discriminator follows the spec's evaluation rule:
  • accuracy_drop >= AccuracyDropThreshold AND same-job hw.gpu.sdc.* counter rose during the job window → kind=vendor_signaled (full confidence)
  • accuracy_drop >= AccuracyDropThreshold * 2 alone → kind=accuracy_only (partial confidence)
  • sub-2x-threshold without vendor signal → no verdict (conservative-by-default noise band)
• False-positive guards (spec §"Edge cases"): dataset-checksum mismatch suppresses verdicts; checkpoint-step mismatch suppresses verdicts. Both guards are opt-in (only fire when both sides set the field).
• Wiring: patterndetectorprocessor projects gen_ai.training.eval_accuracy* + hw.gpu.sdc.* log records, runs the detector, emits the verdict with promoted scalars per issue patterndetectorprocessor: promote operator-facing scalar attrs onto verdict records #270.
• Schema: module/pkg/patterns/testdata/silent_data_corruption_verdict.schema.json pins the verdict wire format; 13 schema-drift falsifiers + conformance round-trip for both branches.
• Docs: spec status flipped to shipped; docs/patterns/README.md table updated; docs/ATTRIBUTES.md registers the 13 new customer-stable attribute keys (hw.gpu.sdc.{delta,kind}, gen_ai.training.eval_accuracy.*, gen_ai.training.eval_set.*, gen_ai.training.checkpoint.*, gen_ai.training.job.{start,end}_unix_nano, tracecore.alert.silent_data_corruption.*) and adds the silent_data_corruption row to the per-pattern matrix.

Verdict remains explicitly advisory per spec §"Detector evaluation rule" — both branches route the operator to a same-recipe + same-seed re-run on different hardware as the disambiguator, because SDC repro is non-deterministic.

Test plan

• cd module && go test ./pkg/patterns/... ./processor/patterndetectorprocessor/... -count=1 — green (19 SDC library tests incl. 13 schema falsifiers, 7 wiring tests).
• cd module && go vet ./... — clean.
• golangci-lint run ./... — 0 issues (via pre-commit hook).
• attribute-namespace-check — 82/82 unique attribute literals documented (via pre-commit hook).
• Full module go build ./... — clean.

Integration gaps (upstream-blocked, documented in-place)

• gen_ai.training.eval_accuracy upstream framework instrumentation (spec §"Open questions" 5) — projection tested against hand-crafted records; real-world input arrives once the OTTL recipe lands.
• hw.gpu.sdc.* OTTL recipe for DCGM SDC catcher / row-remap / AMD ECC counters — blocked on RFC-0014 PR-B (metrics→logs pattern). Cross-vendor support depends on hw.gpu.sdc.* semconv that doesn't exist yet (spec §"Open questions" 2).
• Baseline-accuracy provenance + cross-run state are v0.3+ product questions (spec §"Open questions" 1, 4) — the detector accepts operator-stamped baselines today.

Add pattern #13 (silent data corruption) detector. Surfaces suspected SDC when an eval-accuracy regression (`gen_ai.training.eval_accuracy` vs baseline) crosses the configured threshold; a same-job `hw.gpu.sdc.*` counter rise during the job window flips the verdict to high confidence (`kind=vendor_signaled`). The verdict is advisory — both branches recommend a same-recipe re-run on different hardware as the disambiguator. New config knobs: `sdc_accuracy_drop_threshold` (default `0.005`) and `sdc_accuracy_only_multiplier` (default `2.0`).

[trilamsr]

Reviewer SHIP-WITH-NITS. Tracking yellow finding (threshold-lowering guidance for sub-0.5pp SDC) in #348 for follow-up. Blue nit (panic guard) is defensive-only — library tests pass; punted. Question about 0.5pp citation moves to #348.