Summary

Resolves #614. Netclaw's OpenAI-compatible provider now parses llama.cpp's timings object from chat completion responses, surfacing KV cache hit counts and server-side throughput/latency data that were previously ignored.

What's new

Provider-side parsing (OpenAiCompatibleChatClient.ParseUsage)

• Reads timings.cache_n → UsageDetails.CachedInputTokenCount
• Stores prompt_ms, prompt_per_second, predicted_ms, predicted_per_second in UsageDetails.AdditionalCounts using integer-encoded keys (microseconds / ×100 scale factors, since M.E.AI types AdditionalCounts as AdditionalPropertiesDictionary<long>)
• Graceful degradation: all fields stay null when the provider doesn't emit a timings object (OpenAI, Anthropic, vLLM, TGI)

Protocol surface (UsageOutput, SessionOutputDto)

• New typed fields: PromptMs, PredictedPerSecond (decoded from AdditionalCounts), plus CachedInputTokens/ReasoningTokens which were already in UsageOutput but missing from the SignalR wire DTO
• LlmSessionActor.EmitUsageOutput decodes the scale factors and populates UsageOutput
• SessionOutputDtoMapper maps all new fields in both directions

CLI metrics (HeadlessChannel)

• JSON envelope (chat -p --json) now includes usage.cachedInputTokens, usage.promptMs, usage.predictedPerSecond, ttftMs (client-side time-to-first-delta), and totalMs (client-side prompt→turn-completed wall time)
• Plain-text [usage] line extended with cached=, prompt_ms=, tok_s= so the eval runner can parse metrics without requiring --json mode

Eval harness (evals/run-evals.sh)

• New eval_metrics SQLite table captures per-prompt performance data
• store_metrics() parses the [usage] line after every prompt
• print_metrics_summary() prints per-category and overall performance stats at the end of the run

Verified end-to-end

Docker-isolated 2-turn test against live llama.cpp server (Qwen3.5-27B-UD-Q4_K_XL, dual R9700):

One-iteration eval run on dev branch printed the full performance summary table across 23 prompts with per-category breakdowns.

Code review fixes (applied after initial pass)

• Extracted magic string keys (prompt_us, predicted_tok_per_sec_x100, etc.) as internal const on OpenAiCompatibleChatClient with a cross-reference comment on the consumer side in LlmSessionActor
• Rewrote TryGetLong/TryGetDouble with idiomatic JsonElement.TryGetInt64/TryGetDouble instead of the (value = ...) is var _ trick
• store_metrics uses plain INSERT instead of INSERT OR REPLACE so PK collisions fail loudly
• Removed a Fact(Skip = ...) live-server integration test — the same coverage exists hermetically via StreamingResponse_WithTimings_SurfacesCachedTokens

Test plan

• dotnet build — all projects compile, 0 warnings
• dotnet slopwatch analyze — 0 issues
• dotnet test src/Netclaw.Daemon.Tests (ParseUsage + streaming) — 9/9 pass
• dotnet test src/Netclaw.Cli.Tests — 413/413 pass
• End-to-end Docker test: 2-turn headless run shows all timing fields populated in JSON output
• evals/run-evals.sh with 1 iteration — performance metrics table prints correctly per-category and overall

Out of scope (follow-ups)

• Multi-turn eval cases — PR feat(cli): add --resume and -p flags to netclaw chat for scripted multi-turn sessions #613 added chat -p --resume but multi-turn behavioral/performance cases aren't wired into run_all() yet. Needed to validate KV cache growth across turns, especially for tool-call flows where sub-turn serialization could silently break cache hits.
• Caddy session affinity — round-robin between GPU0/GPU1 currently defeats the prompt cache 50% of the time for same-session follow-ups. The metrics captured here are what's needed to measure the delta when affinity is added.