feat(cloudflare): add telemetry collection and /metrics endpoint

[vahidlazio]

Summary

Add telemetry collection to the Cloudflare resolver with the same metric names as all other providers, enabling shared Grafana dashboards.

Key discovery: scheduler.wait(0) unfreezes CF Workers timers

CF Workers freeze Date.now() and performance.now() during synchronous CPU work (Spectre mitigation), making internal latency measurement appear impossible. We discovered that scheduler.wait(0) — a zero-delay yield to the runtime — unfreezes the clock with no measurable overhead. This enables inline CPU time measurement identical to other providers, without any external API dependency.

What's included

• Inline resolve latency via scheduler.wait(0) + Date.now() (1ms resolution, Grafana interpolates to sub-ms)
• Resolve rate telemetry by reason (MATCH, NO_SEGMENT_MATCH, ERROR, etc.)
• /metrics endpoint serving Prometheus exposition format (confidence_resolve_latency_microseconds histogram + confidence_resolves_total counters)
• Queue-based aggregation — per-isolate deltas accumulated into KV-backed cumulative snapshot
• Backend telemetry — WriteFlagLogsRequest with SDK ID (SDK_ID_CLOUDFLARE_RESOLVER = 25), resolver version, resolve rates, and latency histogram
• Cache-Control: no-store on /metrics response
• Deployer auto-creates KV namespace for metrics storage

Data flow

Hot path:     Date.now() → resolve → scheduler.wait(0) → Date.now() → elapsed_us
              Push {elapsed_us, reasons} to PENDING_METRICS
wait_until:   Drain metrics → TELEMETRY.record_latency_us / mark_resolve
              → checkpoint() → Queue
Queue consumer (batched, all isolates):
  ├─ KV: read cumulative → accumulate deltas → write snapshot + prometheus text
  └─ POST aggregated WriteFlagLogsRequest to Confidence backend
/metrics:     Read prometheus text from KV

What was explored and removed

The initial approach sourced latency from Cloudflare's GraphQL Analytics API (workersInvocationsAdaptive percentiles). This worked but added complexity: cursor-based pagination, rate-limiting, percentile-to-histogram distribution, cached percentile estimation. Once scheduler.wait(0) was proven to unfreeze timers, the entire GraphQL machinery was removed (~360 lines), leaving a clean inline timer identical to other providers.

Test plan

• cargo check -p confidence-cloudflare-resolver compiles clean
• cargo test -p confidence_resolver -- telemetry passes (25 tests)
• Deploy to test account and verify /metrics returns valid Prometheus text
• Verify resolve reasons accumulate correctly (MATCH, NO_SEGMENT_MATCH, ERROR)
• Verify latency histogram matches warm resolve times (~1-2ms p50)
• Verify telemetry appears on Grafana dashboard (resolve rate continuous, latency continuous)
• Verify scheduler.wait(0) adds no measurable latency to resolve path (0ms overhead in 19/20 samples)
• Verify metric names match other providers (confidence_resolve_latency_microseconds, confidence_resolves_total)

🤖 Generated with Claude Code