feat(chart): plot the summary p95 over time, not the trace window

[htcom-code]

Closes #78

Why

The latency chart bucketed the in-memory trace window (500 rows) per minute, so the time it covered shrank with load. At 67 req/s that was 7.5 seconds, one bucket, and during a 4,000-request burst the chart was a single dot. On a fresh connect its first points were the platform's ring-buffer dump, seconds of traffic from before the page connected, drawn under "over time". The p95 tile showed the worst per-module p95 from metrics, a third source that agreed with neither.

What

• Source. The chart plots summary.p95LatencyMs, one point per summary frame received on this page. Each point is stamped with the tick the ready ack stated, or 1 s when no ack arrives. Traces no longer reach the chart: traffic cannot shrink it, and the connect dump does not appear on it.
• Time axis. The x axis spans the selected range and ends now, not at the newest point. When the stream goes quiet the line slides left, so a frozen chart no longer looks live. Before the range fills, the left side stays empty, so the axis shows how much has been collected. The ends are labelled −15m and now.
• Gaps. The line breaks where two neighbouring points are more than 2 × tickMs apart, the same slack the silence watchdog uses. A point alone between two gaps is drawn as a dot, not a ramp.
• Buffer. The series is page-scoped and holds up to six hours whatever range is selected, so widening the range shows what already arrived. It survives a rebuilt stream or a pause; a reload clears it.
• Setting. New pc:chartRange (5m · 15m · 1h · 6h, default 15m). It is in the settings dialog and syncs across tabs through fix(settings): pick up changes made in another tab #76. It uses the TimeRange type, which had no users before.
• Header. It names what a point is: ms · p95 of 60s · ≤200 traces, taken from summary.windowMs and ready.capacity. Anything the platform did not state is left out.
• p95 tile. It reads summary.p95LatencyMs, so it equals the chart's newest point. It shows a dash until the first summary arrives instead of falling back to a different measurement.
• Status mix. The header says last N req: it counts the newest rows in memory, not all traffic. summary has no per-status split, so this stays count-based.
• Sample data. The curve is still labelled sample data. It now carries timestamps (every 15 s over 6 h), so it fills every range, and it is anchored to its own last point rather than the clock.
• Removed. deriveP95 is gone. The dead REST snapshot path now returns an empty series because it has no summary.
• Stale marking. The chart is now dimmed by the summary channel, which feeds it, instead of trace.

Verification

• 20 net new tests (237 total). They drive the real hook with all three captured connect sequences and check the rendered geometry.

• Revert checks: each old behaviour was put back one at a time, and the matching tests failed.

  Put back	Tests that fail
  Right edge pinned to the last point	2
  Line never breaks	4
  Traces feed the series	5 (including the burst and connect-dump scenarios)
  Series reset on each connection	1
  Tile reads metrics	2

• Browser, sample mode on vite: the chart draws on the time axis, the range setting reaches it (−6h, 1,441 points across the full width), and Status mix and the tile read as intended.

• Not checked live: the backend was down, so a live run with gaps, silence and a burst against a real platform is still owed.

• Gates: lint is clean (the same 7 warnings as main), and the build passes.

• Full suite run serially: 237/237 here and 217/217 on main. In a parallel run on a heavily loaded machine (load average ~10), the IndexedDB timing scenarios (retention-limit, trace-window, quota-pressure) fail 2–3 at a time on main as well. This is existing flakiness and not from this change.