fix: stabilize dense live line history

[brandtnewlabs]

Summary

• anchor dense line min/max decimation buckets to absolute time instead of the moving viewport origin
• anchor bounded denoising ranges to the same stable absolute-X grid
• preserve extrema, fixed path order, and the canvas-width output bound
• add rolling-window regressions for both raw and simplified dense paths
• update the drawing-layer and denoising documentation plus changelog

Root cause

The dense-path bucket index was derived from time - winStart. As winStart advanced, historical samples repeatedly crossed pixel-column boundaries, changing each bucket's selected min/max representatives. After fixing that, the optional RDP simplifier still partitioned its input by the first visible point's array index, so its range endpoints shifted whenever a point left the window. Both stages could therefore select different historical vertices.

Decimation buckets now remain fixed in absolute time, and simplification ranges remain fixed on an absolute screen-pixel grid while the time window and layout are unchanged. Fully visible history translates left intact; only boundary ranges change.

Closes #271.

Validation

• npm run verify — 102 suites passed; 1,574 tests passed, 5 skipped
• React Doctor — 100/100, no issues
• agent-device on iPhone 17 Pro simulator — Line, 1m, 20 trades/sec, normal volatility, Trade stream off, Degen off; reviewed separate denoising-off and denoising-1.5 px captures frame-by-frame

[brandtnewlabs]

Agent-device validation video (corrected)

▶ Watch/download the corrected 16-second iPhone 17 Pro capture

Recorded from the fresh Metro bundle after the dense 1-minute window had saturated at 20 trades/sec, with normal volatility, Trade stream off, and Degen off.

• First 8 seconds: denoising off
• Hard cut (no toggle transition shown)
• Final 8 seconds: denoising on at 1.5 px

The axis scale remains unchanged within each segment. Retained interior peaks and valleys translate without reshaping; only the live edge receives new geometry. This replaces the earlier clip that still exposed the moving simplifier-range bug.

[brandtnewlabs]

Performance profile

Profiled the saturated 1-minute Line chart at 20 trades/second with denoising disabled and at 1.5 px. No measurable performance regression was found.

• Process snapshots: 56.0% CPU raw vs 49.2% denoised (directional simulator data).
• Matched 10-second native samples: 832 vs 829 main-thread samples, 778.0 MiB peak for both, and no new path-building hotspot.
• React work: 15.3 vs 15.1 commits/second and 5.7 vs 6.1 ms render time/second; chart/path hooks did not appear in the React re-render list.

Agent-device does not expose Apple frame-health metrics for a simulator, so this does not claim FPS or dropped-frame numbers; that requires a connected physical iOS device.

Full profile report · Native samples: raw / denoised · React profiles: raw / denoised