fix(dashboard): TPS over generation time (duration − TTFT) + reasoning-aware numerator (#13130)

[ahmedhosnypro]

Summary

Fixes the TPS (tokens-per-second) metric confirmed in #13130: dashboards and usage history computed a wall-clock rate (tokens_out / full request duration), which collapses for thinking/interleaved models (kimi k2.6/k2.7/k3) whose queue + prefill + first-token window dwarfs the generation window.

Verification done on a live instance (nvidia/moonshotai/kimi-k3, interleaved thinking + tool calls)

• Streamed a request capped at max_tokens: 3000 → finish_reason: length while visible content was only ~370 tokens (+6.3k chars of reasoning_content). The cap can only be reached if completion_tokens includes thinking tokens, and OmniRoute's recorded tokens_out matched the stream exactly (recorded TPS = 14.4 tok/s = true all-output throughput). So the numerator was already correct for shapes that fold reasoning into completion_tokens.
• What WAS wrong: RequestLoggerV2 + getModelLatencyStats divided by full wall-clock duration, contradicting the repo's own rule in open-sse/utils/generationThroughput.ts ("tok/s MUST exclude TTFT"; the response-header path already follows it).
• Latent gap: nothing added tokens_reasoning into the rate — a provider that excludes reasoning from completion_tokens while reporting completion_tokens_details.reasoning_tokens (the reporter's kimi OAuth endpoints are suspected) would still undercount.

Changes

• src/shared/utils/logTps.ts (new): resolveGenerationMs() (duration − TTFT, full-duration fallback) and resolveTpsOutputTokens() (max(tokens_out, tokens_reasoning) — never double counts, since reasoning is a subset when the provider folds it in).
• call_logs.ttft_ms (migration 176_call_logs_ttft_ms.sql + ensureCallLogsColumns reconciliation): TTFT from streamTiming is threaded through persistAttemptLogs for streaming completions; NULL for non-streaming/legacy rows.
• Call-logs API: summary rows expose ttft; dashboard grid adds an opt-in TTFT column (the requestLogger.columns.ttft i18n key already existed in all 51 locales), the TPS cell tooltip breaks down TTFT / generation window / reasoning tokens, and the request-detail modal shows a TTFT tile with the generation window.
• getModelLatencyStats: per-row TPS sample now uses generation time + the reasoning-aware numerator (feeds auto-combo speed ranking).
• Playground (Chat/Compare client-side TPS): same generation-time formula via the shared helper.

Tests

• New tests/unit/log-tps-13130.test.ts — generation-window math, reasoning-aware numerator (incl. the no-double-count and provider-excludes-reasoning cases), accumulateLatencySample behavior, and ttft_ms persistence round-trip.
• Updated latency-stats-ttft-6875.test.ts and playground-stream-metrics.test.ts (they pinned the old wall-clock formula), db-schema-columns-split.test.ts covers the reconciler.
• node --import tsx/esm --test on all touched suites: green; eslint (with suppressions): clean; npm run typecheck:core: clean.

Refs #13130

---

⚠️ base-red inherited: #12732 — the release/v3.8.51 tip was red when this PR was opened; any suite failures that also fail on base are inherited, not from this change.

[diegosouzapw]

This is the more complete fix for the TPS problem — it doesn't just add the TTFT column
(which #12148 also does, with an identical ALTER TABLE call_logs ADD COLUMN ttft_ms
statement), it actually fixes the root formula bug (wall-clock vs. generation-time) and adds
the reasoning-aware numerator that getModelLatencyStats/auto-combo speed ranking and the
playground also needed. 18/18 tests pass locally. One mechanical blocker: the new migration
176_call_logs_ttft_ms.sql collides with 176_xp_action_counts.sql already on the release
tip — needs renumbering to 179. We're also reconciling this with #12148, which covers the
same DB column/UI surface; expect a note here once that's settled.

[diegosouzapw]

Maintainer note for the merge: this will be squash-merged with Co-authored-by: b3nw <189466+b3nw@users.noreply.github.com> — the TTFT plumbing here (the call_logs.ttft_ms migration, the attemptLogging → chatCore → callLogs threading and the RequestLoggerV2 column) is the same infrastructure @b3nw landed first in #12148. #12148 will be closed once this merges, with credit preserved through that trailer. No action needed from you.

[diegosouzapw]

Migration number coordination: several open PRs claim the same slot, and the release tip is already at 189 (187_proxy_logs_attempts, 188_call_logs_reasoning_encrypted, 189_quota_schedules). This PR's migration (187_call_logs_ttft_ms.sql) gets 190. Please rename the file and update every reference to the number. Slots go by PR age, and merges happen in ascending order so the numbering gate stays contiguous.