QVAC-13715 feat: Profiler Operation Transport + Load/Download Metrics + Stream Profiling

[opaninakuffo]

🎯 What problem does this PR solve?

• No mechanism to surface operation-level metrics (download, cache, checksum, initialization timing) to client
• No per-call profiling control (RPCOptions) for loadModel/downloadAsset
• Profiler export formatted all gauges as durations (misleading for bytes/throughput)
• No test coverage for profiler transport contract
• Stream profiling lacked server-side metrics
• Negative clientOverhead values possible due to clock skew

📝 How does it solve it?

Operation Event Transport

• Add OPERATION_EVENT_KEY + operationEventSchema in profiling schema
• Inject operation events into __profiling envelope (survives response parsing)
• Client profiler records serverMeta.operation events

Delegated loadModel Propagation

• Extend ResponseWithDelegation to carry OPERATION_EVENT_KEY
• Extract/attach provider operation events in both unary and streaming paths
• Forward operation metadata to final consumer response

Stream Profiling (Trailer Frame)

• Server emits profiling-only trailer frame at stream end with __profilingTrailer marker
• Client extracts metadata then skips trailer (no yield to user code)
• Server profiler caches delegation/operation from stream chunks, emits in trailer
• Delegated streams attach breakdown per-chunk (profiled path only, zero overhead on non-profiled)
• Conditional serverMeta assignment preserves metadata if later chunks lack it
• clientOverhead clamped to Math.max(0, ...) to prevent negative values

Load/Download Metrics

• Add DownloadStats, ResolveResult, DownloadMetricsHooks types
• Add resolveModelPathWithStats() with base/stats mode
• Add download-stats.ts wrappers (HTTP/Registry/Hyperdrive) with:
  • Progress-based bytes tracking
  • Download timing/speed derivation
  • Cache-hit/miss hooks
  • Checksum timing accumulation

Handler Operation Events

• loadModel gauges: totalLoadTime, modelInitializationTime, downloadTime, totalBytesDownloaded, downloadSpeedBps, checksumValidationTime
• loadModel tags: modelType, sourceType, cacheHit
• downloadAsset gauges: totalDownloadTime, downloadTime, totalBytesDownloaded, downloadSpeedBps, checksumValidationTime
• downloadAsset tags: sourceType, cacheHit

Client API

• Add optional rpcOptions?: RPCOptions to loadModel() and downloadAsset()
• Forward through unary and streaming RPC paths

Export Improvements

• Aggregate gauge stats (not just timing)
• Format table by metric type (duration/bytes/throughput/number)

How it was tested

• bun run typecheck
• bun test test/unit/profiler.test.ts
• bun test test/unit/profiler-operation-transport.test.ts
• bun run examples/profiling/basic.ts (manual verification, results below)

This PR depends on #836

[opaninakuffo]

/review

[github-actions]

Tier-based Approval Status

**PR Tier:** TIER1

**Current Status:** ✅ APPROVED

**Requirements:**
- 1 Team Member approval ✅ (1/1)
- 1 Team Lead OR Management approval ✅ (1/1)



---
*This comment is automatically updated when reviews change.*