docs: decoupled fetch/compute canary confirmation (Stage C gate CLEAR)

[WilfordGrimley]

Summary

Lands the finished decoupled fetch/compute confirming canary report + one troubleshooting entry. Report-only (no code); the code it confirms (#228 design → #237 implementation) is already on master. This is the report that was stranded on its worktree branch after the prior session was ejected.

• docs/reports/2026-07-20-decoupled-canary-confirm.md (new, 239 lines)
• docs/troubleshooting.md (+40 lines): the benign pg_type_typname_nsp_index UniqueViolation startup race, confirmed self-healing.

Result — GATE CLEAR

First real at-scale confirmation the #237 decoupling fix worked (400-card prod canary, run_id stagec-canary-decoupled-20260720T235127Z, real write, exit 0):

metric	bundled (prior)	decoupled (this run)
parallel efficiency (400-card / steady-state)	63.1%	95.2% / 99.1%
projected full-pool wall-clock (217,828 cards)	~15.7–16.0h	~10.5–10.9h
extraction failure rate	—	1.5% (< 5% ceiling)
memory peak	~904 MiB	~1.43 GiB (« 24 GB)

Per-card --profile confirmation: corr(extraction_ms, wall_ms)=0.9998, corr(fetch_ms, wall_ms)=-0.249 — fetch is now decoupled from compute wall-clock, not just an aggregate-throughput inference. Index-not-store posture re-verified (only raw bytes cross the fetch/compute boundary; profile JSONL holds ids + ms floats only).

All 5 gate conditions clear. Phase 2 (20k cohort + Stage D dry-run) was NOT run — that is the unblocked next step, still gated on a fresh explicit owner GO.

Verification

• Real write verified: ImageEvidence.objects.filter(run_id=...).count() → 400; spot-checked row has all new signal fields non-null.
• Live API 200 before/after; all 5 persistent containers stayed Up, no restart.
• Fast-forward-clean onto current master (docs-only, no conflict).

Open items

1. WORKERS.md worktree-coordination gap (confirmed a 2nd time this run) — unresolved, flagged for the next long-running worktree dispatch.