docs(oura): recording is motion-gated, and 0x50 MET tracks a varying input

[pipiche38]

Documents three findings about the Oura Gen 3 wire streams from a ground-truth session on 2026-08-02, and records why an earlier-looking correlation on 0x7E is an artifact rather than a result.

What this adds

Recording is motion-gated. The ring does not log on a fixed wall-clock cadence — it stops emitting when the wearer stops moving. On an 86-min flat 6.5 km walk containing a deliberate ~3-min standing stop, 0x7E/0x7F (174 s), 0x47 (150 s) and 0x50 (240 s) went silent together across the stop and resumed at their normal cadence. Three independent streams agreeing rules out a decode drop or a drain artifact.

The consequence matters to anything consuming these streams: a gap must be read as "no motion", not as missing data to interpolate over, and a "samples × cadence" product is not a wall-clock duration. It also re-explains the 0x50 logging holes noted earlier in §6.13 as quiet periods rather than dropped records.

0x47 low_intensity / motion_seconds saturate. Both cap at 29 within a 30 s record; across 168 records of continuous 4.6 km·h⁻¹ walking, 65 % and 82 % of samples sit at ≥ 27. They count seconds containing motion, not an amount or rate of motion — so a step count or cadence cannot be recovered from 0x47 at any scale factor. This is a structural reason independent of the ground-truth test already recorded, and it applies to a Tier-A stream, so a better decode will not fix it.

0x50 MET tracks a varying input. Against the deliberate stop, MET falls to a resting floor when the feet stop and recovers when they restart: r = 0.860 over 88 minutes, and — the part that matters — robust: dropping the 3 lowest-activity samples leaves r = 0.777, where the 0x7E field collapses 0.912 → 0.256 under the same test.

This is the transition evidence CLAUDE.md asks for ("prove the method tracks a varying input"), as distinct from the earlier level match which could not exclude a coincidental fit. Two caveats are kept explicit in the text: MET lags ~1 min on recovery, and this validates that MET tracks intensity, not that its absolute scale is calibrated. It stays Tier B and unscored.

Why the 0x7E r = 0.912 is not a result. Recorded so it is not rediscovered later and mistaken for validation: the input barely varied (98 % of samples effectively identical), and the one interval with real contrast produced no rows at all because recording is motion-gated. A useful step-decoder test needs cadence variation — mixed slow/fast walking, stairs, walk/jog intervals — not a longer metronomic walk.

Verification

• Documentation only. No code, no stored data, no test changes — the diff is docs/OURA_PROTOCOL.md, +87 / −0.
• Branched directly off main, so it is purely additive and does not touch 73c1dd8 (Say what the Oura SpO2 unit tag actually means, and correct the mapping table (#1043) #1045) or any other v9.3.1 work.
• Findings come from a live Gen 3 ring paired to NOOP, with per-second cadence ground truth from a Suunto .fit for the same walk; every figure quoted is from that session's decoded streams.
• No CI job covers docs/; swift-packages is unaffected as no package source changes.

[pipiche38]

@ryanbr would appreciate if you could review and merge it if possible ?