feat(codemode): show eval tool-call throughput

[code-yeongyu]

Summary

• show exact nested tool-call count and calls/second in completed eval TUI headers
• use true wall-clock elapsed time for both the visible duration and throughput denominator
• preserve kernel-reported duration and the existing senpi.eval.execution event contract
• cover zero-duration, partial/error, multi-cell, legacy, and 40/80/120-column cases
• add hermetic real-source PTY/xterm/Chrome QA

Test evidence

• RED: renderer assertions saw ... 1.00 calls/s · 1s ... when wall time was 2s; runtime details lacked toolCallCount
• GREEN: focused renderer suites 58/58; runtime metadata suites 8/8
• Full @code-yeongyu/senpi-codemode suite: 565 passed, 6 skipped
• npm run check: passed
• npm run build: passed

Manual QA

Command:

node .agents/skills/senpi-qa/scripts/scenarios/eval-throughput-badge-qa.mjs --self-test --evidence eval-throughput-badge

Result: 9/9 passed.

Observed real TUI header:

╭─ eval js done ✓ · 2 calls · 25.32 calls/s · <1s

Evidence includes raw PTY bytes, xterm.js grid JSON/HTML, Chrome PNG, sanitized fake-provider requests, and cleanup/auth receipts under local-ignore/qa-evidence/20260817-eval-throughput-badge/.

Dual visual QA verdicts: PASS / PASS.

---

Summary by cubic

Shows eval tool-call throughput in the TUI header using wall-clock time. Previously the header showed only kernel duration; now it displays “N calls · X.XX calls/s · elapsed”, making throughput visible while preserving the senpi.eval.execution event contract.

• Rendering: Badge appears only for final, single-cell, completed evals; hidden for partial/running/error and multi-cell results. Pluralization is correct; zero elapsed shows “0.00 calls/s”; positive calls without positive wall time show “n/a calls/s”. The final metadata line now uses wall time for “took …” and also includes calls and calls/s; timeout/reset badges remain after elapsed.
• Data model: EvalToolDetails adds wallDurationMs and toolCallCount (additive). Kernel durationMs is unchanged; the renderer prefers wallDurationMs for visible elapsed and throughput.
• Docs and tests: packages/senpi-codemode/CHANGELOG.md and packages/senpi-codemode/changes.md document behavior and expected conflict zones. Tests updated; added hermetic PTY/xterm/Chrome QA scenario at .agents/skills/senpi-qa/scripts/scenarios/eval-throughput-badge-qa.mjs. No migration required.

^{Written for commit 3908a8c. Summary will update on new commits.}