Recipe extension: emit hw.gpu.nvlink.* + wire metrics path for pattern-1 detector

[trilamsr]

Blocker

Pattern #1 (NVLink silent degradation) is the next NORTHSTAR detector on the
v0.3.0 ladder, but the input plumbing it requires is not yet shipped. The
detector cannot be built until this lands.

Source-of-truth: docs/patterns/pattern-1-nvlink-degradation.md.
That doc specifies the input signal as the customer-stable
hw.gpu.nvlink.io Counter with attributes:

Attribute	Value
hw.gpu.nvlink.link	0 .. N-1 (18 on H100, 12 on A100)
network.io.direction	transmit / receive

Resource attributes: hw.id, hw.gpu.index.

What's missing

1. OTTL transform — raw DCGM → customer-stable semconv. The recipe at
   docs/integrations/prometheus-scrape.md
   scrapes dcgm-exporter but only stamps gpu.vendor. It does NOT rename
   DCGM_FI_PROF_NVLINK_L{0..17}_{TX,RX}_BYTES into the
   hw.gpu.nvlink.io Counter shape with per-link / per-direction attrs. The
   semconv itself is PROPOSED in
   docs/proposals/semconv-hw-gpu-extensions.md
   §3 and docs/rfcs/0005-dcgm-receiver-scope.md, but no recipe currently
   emits it.

2. Metrics-side detector plumbing. patterndetectorprocessor is logs-only
   today (processor.WithLogs(...) in factory.go — no WithMetrics). The
   NVLink signal is a Counter, not a log record. Pattern ci(deps): bump the gh-actions group with 5 updates #1 needs either:

   • a metrics input path on patterndetectorprocessor, OR
   • a metrics→logs converter (rule-engine style — emit one
     nvlink_degraded log per offending link, analogous to the PromQL alert
     rule in the pattern doc).

Why this matters now

Per feedback_adopt_over_build.md: don't fake attribute projections. The
upstream dcgm-exporter emits raw DCGM_FI_PROF_NVLINK_L*_TX_BYTES /
*_RX_BYTES (18 links × 2 directions × N GPUs per node). Stamping those as
hw.gpu.nvlink.io with the customer-stable attrs is a recipe-shaped task
— transformprocessor OTTL is the right primitive, and the work belongs
in docs/integrations/prometheus-scrape.md (or a sibling
docs/integrations/dcgm-nvlink-recipe.md) rather than inside the
detector.

Proposed plan (separate PR, blocks detector PR)

PR A — recipe extension (this issue):

1. Extend the OTTL transform/gpu_vendor block (or add transform/dcgm_nvlink)
   to:
   • For each DCGM_FI_PROF_NVLINK_L{N}_TX_BYTES series: emit
     hw.gpu.nvlink.io Counter with hw.gpu.nvlink.link=N,
     network.io.direction=transmit.
   • For each DCGM_FI_PROF_NVLINK_L{N}_RX_BYTES: same with
     network.io.direction=receive.
   • Map gpu_uuid / UUID label → resource attr hw.id.
   • Map gpu / GPU label → resource attr hw.gpu.index.
2. Decision needed: metrics input path on patterndetectorprocessor, OR a
   thin metrics→logs converter (OTTL transform/metrics → log signal)?
   Either way: pattern-detection consumes the customer-stable attribute
   namespace, never the raw DCGM_* series.
3. Document in
   docs/integrations/prometheus-scrape.md
   alongside the existing gpu.vendor mapping table.

PR B — NVLink detector (blocked on PR A):

• module/pkg/patterns/nvlink_degradation.go + _test.go mirroring
  xid_correlation's two-layer shape.
• appendNVLinkDegradationVerdict in
  module/processor/patterndetectorprocessor/.
• PatternIDNVLink = "1", EvidenceKindNVLink = "nvlink".

Out of scope

• The semconv proposal merge — docs/proposals/semconv-hw-gpu-extensions.md
  remains PROPOSED until upstream OTel semconv adopts it. Tracecore's
  customer-stable contract (RFC-0013 §3) is the local pin while upstream
  catches up.
• Blackwell-class NVSwitch SXid evidence path — pattern ci(deps): bump the gh-actions group with 5 updates #1 v0 scope is
  Hopper/Ampere per MILESTONES.md M17.

Pre-existing references

• Pattern doc: docs/patterns/pattern-1-nvlink-degradation.md
• Semconv proposal: docs/proposals/semconv-hw-gpu-extensions.md §3
• DCGM receiver RFC (deleted, but signal table still load-bearing):
  docs/rfcs/0005-dcgm-receiver-scope.md §hw.gpu.nvlink.io
• M17 milestone definition: MILESTONES.md (search "M17")
• Pivot RFC adoption matrix: docs/rfcs/0013-distro-first-pivot.md §2

[trilamsr]

Recipe-half of #260 closed by commit adbba39. The DCGM hw.gpu.nvlink.io OTTL transform landed in commit 0baa557 (already extends transform/dcgm_to_hw_semconv with per-link/per-direction attribute lift). Detector PR-B (processor.WithMetrics extension) tracked under RFC-0014.