ConnectionTrace session duration rounds half-ties differently on Swift and Android #87

Finding

ConnectionTrace.sessionHeldSuffix/1 is a declared Swift/Kotlin byte contract, but native formatting disagrees at millisecond values that land on a tenth-second half-tie.

Machine-readable native reproduction:

The fresh differential pilot therefore fails session_held_half_tie at 50 ms. This is not harness normalization noise: Swift calls Foundation String(format: "%.1f", ...); Android calls Java/Kotlin "%.1f".format(...), and the native runtimes select different half-tie results.

Reachability and impact

The helper is public on both platforms and is used by both live disconnect emitters:

• Swift BLEManager appends it to connect down (uptime ends...) while Connection Test Mode is active.
• Android WhoopBleClient appends it to the corresponding live log line.

This changes diagnostic text by 0.1 seconds only; it does not affect reconnect behavior, storage, or user health data. Severity is bounded P2 parity contract / P3-low operational. Exact half-ties are possible because both callsites derive integer milliseconds, although typical sessions last much longer.

RED / acceptance

Add mirrored public tests for at least 50, 150, and 250 ms and require byte-identical output under an explicit canonical rounding policy. Preserve existing contracts:

• negative millis => empty suffix (unknown session start)
• zero => after 0.0s
• ordinary 432/6800/120000 ms outputs unchanged
• locale remains fixed to a decimal point

The parity harness must then pass the complete curated Readout corpus without an exemption or output normalization.

Provenance / routing

The helper originated in upstream PR ryanbr#1023 for issue ryanbr#1020. That PR tested ordinary tenth-second values but not half-ties. Duplicate searches across open/closed issues and PRs in both ryanbr/noop and bhelm/noop found no existing rounding report or fix.

Fix routing: upstream-only production PR against ryanbr/noop; do not patch production only in the fork.