feat: v0.3.7 — idle drift cruise (ambient field baseline)

[Jess Sullivan (Jesssullivan)]

Root cause

driftAngle/driftSpeed have been initialized per blob since the pre-Phase-A baseline but were never read by the physics loop — dead code. With no pointer, scroll, or device-motion input, idle motion was only a zero-mean jitter random walk (territorial jitter, neutral drift, brownian, escape kicks) that the per-substep *= 0.992 damping erases with a ~1.4s half-life: blobs jiggled near their territory instead of drifting. The only sustained forcing was gravity from devicemotion, which desktops never emit. Separately, the component rAF clamp of 0.033s pre-empted the 0.3.6 fixed-timestep accumulator: sustained rendering below 30fps dilated simulated time proportionally (~3x slower at 11fps).

Fix

• updateScreensaverPhysics now steers each blob's velocity toward its own driftAngle heading at driftSpeed * DRIFT_CRUISE_SPEED_SCALE every substep (DRIFT_CRUISE_STEERING = 0.05, scale 3). The settled cruise (~1.5-3.9 units/s in the -40..140 physics space) is the "idle blobs drift" / "ambient field: always on" baseline of docs/physics-feel-contract.md; the existing ~8s re-heading roll and wall-bounce re-randomization keep paths wandering, and pointer/scroll/gravity fields still bias motion on top exactly as before.
• rAF frame-delta clamp raised 0.033 -> 8/60 to match BlobPhysics.tick()'s own catch-up ceiling (8 substeps of 1/60s); sustained rendering down to ~7.5fps now keeps real-time motion speed, tab-resume jumps still bounded by the engine substep cap.

How verified

• New tests/unit/blob-physics-idle-drift.test.ts (seeded xorshift32, same pattern as the timestep test): idle mean net displacement rises from ~4.3-5.0 to ~16.4-25.5 units over 12 simulated seconds vs a driftSpeed-zeroed baseline; speeds and positions stay bounded. Full suite: 214 passed.
• 300-simulated-second bounds run: positions stay inside the physics walls, peak speed ~0.15 units/substep, 346 wall bounces across 8 blobs (blobs now actually reach and bounce off walls).
• Headed-Chrome A/B probe of the built package (12s, bbox-center waypoints at 1Hz, 60fps): idle net/path coherence rises from ~0.50-0.64 (decorrelating jitter) to ~0.72-0.81 (directed travel) with mean net displacement ~20 units/12s in the 133-unit viewBox.
• check:release-metadata, svelte-check (0 errors), build, check:package lint, bundle size 11.79 KiB gzip (12 KiB gate; +0.06 over 0.3.6).

Reduced motion

prefers-reduced-motion: reduce behavior is untouched: the component still renders the static frame and never starts the rAF loop. Headed-Chrome probe with a reducedMotion: 'reduce' context confirms 0 of 15 paths move.