feat(cache-proxy): per-request phase timing, in-flight gauge, pprof

[EDsCODE]

Problem

Production investigation of slow cold scans found requests spending 100-400ms each inside the cache proxy, served strictly one at a time, with every externally observable component healthy — no errors, no retries, idle CPU, fast peers. The proxy exposes no per-request timing, no queue-depth signal, and no pprof, so the remaining decomposition (origin fetch vs disk vs response write vs queueing) is unmeasurable. Separately, a continuous stream of co-tenant HEAD requests was found traversing the uncached forward path to origin with no metric counting it.

Change

Instrumentation only — no request-path behavior changes:

• Served. log line gains dur_ms, peer_ms, s3_ms, write_ms (existing fields unchanged, in order). Forward-proxy served. gains dur_ms.
• New cache_proxy_request_duration_seconds histogram, labels {path=block|forward, source}, 1ms–8s exponential buckets.
• New cache_proxy_inflight_requests gauge around HandleProxy — the queue-depth signal. Note: CONNECT tunnels count for their lifetime.
• New cache_proxy_forward_requests_total{method} counter — quantifies the forward-path HEAD stream.
• pprof handlers registered explicitly on the health mux (side-effect import only covers DefaultServeMux).
• docs/metrics.md documents the new family, distinguishing it from the worker-side duckgres_worker_cache_proxy_* view.

Accounting nuance: origin-fetch time under singleflight is attributed to the winning request; followers' waits appear in dur_ms but not s3_ms.

Testing

• go test ./cmd/cache-proxy -count=1 green; -race green except the pre-existing TestHandleConnectLogsOpenAndClose race (fails identically on the base branch).
• New test asserts the in-flight gauge returns to its pre-request value after a served request and the duration histogram records exactly one sample, using the repo's existing dto.Metric helper convention.

Notes

Stacked on #1042 (seek fix) — both touch the serve loop. Merge that first; this PR's base is set accordingly.

🤖 Generated with Claude Code