perf(drag): model dynamics on 7 recorded human drags

[wmehanna]

Summary

Replaces the icon-drag's guessed velocity model (symmetric smoothstep, derived from one recording) with dynamics measured from 7 real human drags, fixing the last flaky test.

What the recording showed

Captured with helpers/event-logger.swift (now committed as a repo tool):

• Velocity is not symmetric: gentle start, a ~2.3x burst around 30-40% of the movement, then a long tapering landing where the final 15% of time covers almost no distance.
• Path is bowed, not straight (~3.3px mean deviation from the straight line).
• Humans pause still ~135ms at the destination before releasing.

Changes

• Baked the measured median displacement curve as HUMAN_CURVE / humanDragCurve in both drag paths (jxa_icon_drag_body in common.sh, dragSegment in gestures.swift).
• Path now arcs with a deterministic sin-bow (side by origin parity) instead of a straight line.
• Added a still landing settle on the exact drop point before release — gives SpringBoard a stationary target, which is where reorder placement was slipping.
• Committed the event recorder for future re-measurement.

Testing

• icon-drag no-op tour: 10/10 (was ~5/6)
• Live suite --bench 3: 42/42, icon-drag 3/3
• Per-drag timing variance collapsed from 7.7-15.5s to 7.55-7.59s (independent consistency signal)
• Curve verified monotonic with exact endpoints
• bats 179/179, shellcheck clean, phone layout restored