[rc1-prep] OTTL recipe: project DCGM_FI_DEV_FB_USED/FB_FREE → hw.gpu.memory.{free,total} log shape (pattern #10 wiring)

[trilamsr]

The CUDA OOM detector (#303, landed in chore/v1-rc1-knowledge-gaps wave) consumes hw.gpu.memory.{free,total} log records but the prometheus-scrape recipe does not yet project these from DCGM_FI_DEV_FB_USED / DCGM_FI_DEV_FB_FREE. Until this lands, the detector cannot fire on a real install.

Cross-ref:

• docs/patterns/10-cuda-oom-deceptive.md §Signal sources
• module/processor/patterndetectorprocessor/cuda_oom.go
• Sibling to OTTL recipe: derive hw.errors.delta log shape from hw.errors counter (pattern #3 wiring) #273 (pattern Add NORTHSTARS, RFC-0002, and Q1 MILESTONES #3 wiring), OTTL recipe: derive hw.gpu.throttle.duration.delta log shape from hw.gpu.throttle.duration counter (pattern #4 wiring) #282 (Wave 1: governance bootstrap (CODEOWNERS, DCO, signing) #4), OTTL recipe: derive tracecore.alert.pcie_rate_collapse.* from hw.gpu.io counter (pattern #5 wiring) #284 (Apply Wave-1 follow-ups: tests, automation, doc-truth #5)
• Per RFC-0014 metrics-to-logs bridge contract

Add OTTL stanza to docs/integrations/prometheus-scrape.md (or a new dcgm-extension recipe) that projects FB used+free as log records the detector library consumes.