P2: Smaller review findings from the 10.1.0-10.1.1 upstream range (collective) #32

Collective issue for the smaller findings from the independent review of the integrated fork
state (2026-08-18). All arrived with the upstream range d4b7cc8..f2271cd. Each was verified,
but with less depth than the individually filed ones — worth a check before fixing. Filed here
first; forwarding upstream is a separate decision.

• BreathProtocol.kt throws where Swift clamps — the Kotlin protocol player raises on
  out-of-range input that the Swift twin clamps; a malformed catalog entry would crash Android
  and degrade gracefully on Apple platforms.
• hrOnlyFallback misreports under its compile-time gate — the reported fallback state
  does not match actual behavior when the compile-time flag disables the path.
• marginToSigma guard fails at band ≤ 1.5 — the guard band arithmetic breaks down for
  narrow bands.
• StressView / StressScreen re-read 30 days of raw streams on every refresh — no
  caching layer between the screens and the stores; noticeable battery/latency cost on the
  stress screens.
• BreatheScreen uses a 50 ms polling loop — Android breathe animation polls instead of
  deriving state from frame time.

If any of these graduates into real work, it should get its own issue with anchors and a
failure scenario.