fix(dashboard): derive accepted/drafted totals from the per-depth arrays

[dajiaohuang]

Summary

The Per-depth acceptance subtitle and the Drafted / verify call tile read accepted_drafts and drafted_tokens, which _metrics_envelope does not emit — it emits accepted_by_depth and drafted_by_depth. Both surfaces therefore rendered — accepted of — drafted and 0.00 tok/call while the bars beside them drew real per-depth numbers.

Both now total the per-depth arrays they already receive, through a shared depthTotal helper. The helper still prefers a flat total, so nothing changes if the envelope later carries one.

This is the aggregate half of #401. The Recent requests WHEN column is deliberately untouched: no metrics row persists a completion timestamp (the last_metrics entries carry durations only), so a relative time needs a new field on the envelope and on its request-log row — a server-side contract change rather than a dashboard one.

Verification

• bun run typecheck (tsc --noEmit): clean.
• bun run build (tsc -b && vite build): success. The rebuilt mtplx/dashboard/_static is included here; the CSS asset hash is unchanged (index-ZaQvA7xU.css), so only the JS chunk moved.
• The rebuilt bundle was confirmed inside the wheel from python -m build: it carries assets/index-CUmjZF08.js and no longer carries the stale index-CjL8TS30.js.
• The derivation was exercised directly against real envelope shapes with bun: per-depth arrays only gives 13 verify calls · 23 accepted of 36 drafted and 2.77 tok/call; an emitted flat total still wins; empty or absent arrays give — and 0 drafted · 0 verifies.
• python -m build: success (mtplx-2.11.2).
• python -m pytest tests/test_no_mlx_imports.py tests/test_runtime_kpis.py: 13 passed, 8 failed, and all 8 failures are Windows-only artifacts (no mlx; mtplx/thermal.py imports pwd, which does not exist on Windows). The diff changes no .py file. tests/test_public_cli.py cannot be collected on Windows at all (os.geteuid), and CONTRIBUTING's scripts/fresh_venv_smoke.sh expects a POSIX venv layout ($VENV/bin/python), so neither ran here.
• git diff --check: clean.

Benchmark Evidence

Not applicable — this changes dashboard rendering only, not the runtime path.