Generation timing metrics in /stats

Context

Issue #4252 asks /stats to show generation timing separately from session wall time and end-to-end API latency. The low-level timing already exists:

• LoggingContentGenerator measures ttftMs from request dispatch to the first user-visible streamed chunk.
• endLLMRequestSpan derives sampling_ms and output_tokens_per_second.
• ApiResponseEvent already carries request duration, model, prompt id, and output-token count into UiTelemetryService.

The missing link is making the existing TTFT value available to the content-free session metrics used by /stats.

Scope

This change adds live, session-scoped generation metrics to:

• the interactive /stats Session tab;
• the non-interactive /stats text response.

It does not add a second timer, persist timing in daily/monthly token-usage files, change exports, or change the daemon/Web Shell stats schema.

Data flow

LoggingContentGenerator.loggingStreamWrapper
  -> ApiResponseEvent(ttft_ms)
  -> logApiResponse
  -> UiTelemetryService
  -> SessionMetrics.generation
  -> SessionContext
  -> /stats

ttft_ms is optional. Non-streaming responses and streams that finish without user-visible content keep the current behavior and do not create a generation sample.

Metrics and semantics

For each successful streamed response with TTFT:

• TTFT is the existing ttftMs measurement.
• Generation time is max(0, duration_ms - ttft_ms), measured from the first user-visible streamed content through completion.
• TPS is output_token_count / generation_time_seconds. It is unavailable when generation time is zero.

SessionMetrics.generation is created lazily and contains:

• the latest completed request's model, TTFT, generation time, and output-token count;
• total timed request count and TTFT, plus generation time and output tokens for throughput-eligible requests.

The session average TTFT is the arithmetic mean across timed requests. Session TPS is weighted throughput: total output tokens divided by total generation time. Requests with zero generation time contribute to TTFT statistics but not to either side of the session TPS calculation. This avoids division by zero and over-weighting short requests.

Internal helper prompts are excluded from generation metrics. They are not recorded in the resumable transcript, and including them would both surprise users and make live and resumed session values disagree. Main-conversation and subagent requests remain included, matching the existing session-level model statistics.

Compatibility

• ApiResponseEvent.ttft_ms and SessionMetrics.generation are additive and optional.
• Existing recorded events and callers remain valid.
• Existing daily/monthly records continue to contain token and API-duration data only, preserving the ownership boundary documented in issue-4479-token-usage-stats-coordination.md.
• The Session context clone/equality logic copies and compares the optional generation object so the interactive dashboard updates on every completed timed request.

Validation

• Core tests prove aggregation, internal-prompt exclusion, zero-generation handling, session isolation, and reset behavior.
• LoggingContentGenerator tests prove the captured TTFT reaches ApiResponseEvent and remains absent for non-visible streams.
• CLI tests prove non-interactive output and interactive Session-tab rendering.
• i18n tests cover every built-in locale for the new high-visibility labels.