Swift SleepStager gravity fingerprint aliases distinct axes in detect, V1, and V2 caches #66

Summary

The Swift Sleep stager cache fingerprint quantizes gravity as Int((x + y + z) * 1024). Distinct axis vectors with the same sum therefore share a cache key even when the staging recipes observe different motion. Android fingerprints the axes independently and does not reproduce the alias.

Affected public entry points:

• SleepStager.detectSleep/10
• SleepStager.stageSession/6
• SleepStagerV2.stageSession/6

Native RED/control evidence

Tested at fork commit 6df88d0d5d3398fcb4fec62c6cb3b23da3380dae with isolated, env-gated native probes and cache-safe A→B→A sequences. Controls force uncached computation (detect trace sink; stage controls use distinct fingerprints without a cache reset hook).

Swift (SleepStagerDriftProbeTests): all three cache probes RED.

• detect: uncached B = []; cached B incorrectly returned cached A session [1749520800,1749526199].
• V1: uncached/control B shape = 0:180:light, 180:1200:wake; cached B incorrectly returned A shape 0:360:deep, 360:1200:light.
• V2: uncached/control B shape = 0:1610:wake, 1610:1790:light, 1790:1800:deep; cached B incorrectly returned A shape 0:1790:light, 1790:1800:deep.

The mirrored Kotlin probes are GREEN (4/4 full probe class, including these three), BUILD SUCCESSFUL. A and B use gravity rows with equal x+y+z; V2 uses sparse (1,1,-1) samples against (0,0,1) so the nightly jerk floor remains a valid control.

Acceptance

• Swift cache keys distinguish behaviorally distinct x/y/z gravity streams for detect, V1 stage, and V2 stage.
• The isolated A→B→A probes pass on Swift and remain green on Kotlin.
• Unchanged-input memo hits and bounded-cache behavior remain covered.
• No cache reset test hook is introduced into the public production API.
• Differential rollout for the three affected entry points remains halted until resolved.

This is separate from #43 (the quarantined V2 pad-boundary decision).