fix(dashboard): raise p95 sample threshold to 10 turns with p50 fallback

[nicolotognoni]

Summary

• PR-feat(observability): emit patter.cost.* and patter.latency.* OTel span attributes #82's lowered threshold (5 → 2 turns) gave parity with the call-list column but produced a misleading headline on short calls. Live test: n=5 turns → p95=1977ms while p50=309ms, because at n<10 the 95th-percentile collapses to "slowest single turn" and stops being a tail estimate.
• Raised threshold back to 10 across every dashboard surface that exposes a percentile, and added a p50 fallback so short calls still show a useful number (just labelled honestly).
• App-level "Avg latency p95" card now also gates on ≥3 qualifying calls, otherwise it renders "—" instead of an average that's effectively dominated by whichever short call happens to be in the bucket.

Implementation

• MIN_TURNS_FOR_PERCENTILES = 10 exported from LatencyPanel.tsx and MetricsPanel.tsx; MIN_TURNS_FOR_P95_COLUMN = 10 from CallTable.tsx; MIN_TURNS_FOR_AVG_P95 = 10 + MIN_CALLS_FOR_AVG_P95 = 3 from App.tsx. Single source per surface so the threshold can't drift.
• Detail pane (LatencyPanel, both latency view branches of MetricsPanel): below the threshold the p95 box renders p50 (n<10) instead, with a tooltip and footer line spelling out the rule. Realtime calls (single-bucket waterfall) and pipeline calls (stt/llm/tts waterfall) both covered.
• Call list (CallTable.tsx): column renamed "p95 latency" → "Latency" (since it now reports either statistic); rows with turnCount < 10 show <ms> (p50); column header has a tooltip explaining the fallback.
• Sparkline tooltip (Metric.tsx bucketHeadline): when no call in the bucket has ≥10 turns, the headline reads AVG LATENCY n/a (n<10 turns) rather than a fake "0 ms".
• App headline card (App.tsx): avgP95() filters out calls with <10 turns, requires ≥3 qualifying calls, returns 0 otherwise. The card then renders "—" instead of "0 ms" so the empty state is unambiguous.
• New test file dashboard-app/src/App.test.ts (8 cases) covers avgP95 gating and bucketHeadline fallback.
• Re-built the SPA (vite build) and re-synced the inlined bundle into libraries/{typescript,python}/.../dashboard/ui.html via dashboard-app/scripts/sync.mjs so both SDKs ship the updated UI.

Breaking change?

No. The thresholds are internal to the dashboard SPA; SDK API surface is untouched. Users with long calls (≥10 turns) see no change; users with short calls now see honest p50 numbers instead of a noisy p95.

Test plan

• dashboard-app: npm test -- --run → 16/16 pass (8 new + 8 existing)
• dashboard-app: npm run lint (tsc --noEmit) clean
• dashboard-app: npm run build succeeds, 208 kB bundle synced
• libraries/typescript: npm test -- --run → 1516/1516 pass
• libraries/typescript: npm run lint clean
• Manual: open dashboard on a short (n<10) call and verify the detail pane shows p50 (n<10) with a tooltip
• Manual: open dashboard on a long (n≥10) call and verify p95 surfaces as before

Docs updates

• N/A — internal dashboard rendering change. CHANGELOG.md entry added under ## 0.6.1 (2026-05-12).