diag(oura): raw undecoded history-notification capture → oura-raw-<id…>.jsonl

[pipiche38]

Thanks for approving this PT which will help me to diagnose new tage (such as Spo2) on iOS

Adds a RAW capture sidecar complementing the decoded MET/motion dumps: it records the undecoded history-drain TLV notification bytes exactly as received, so after a full connect we know precisely which records arrived. A hole in a decoded file can now be pinned as a decode drop (present in raw, absent in decoded) vs ring-side (absent in both), instead of being ambiguous.

Motivation: an iOS 9.2.1 field export (real Oura ring, Diagnostics capture on) carried oura-activity.jsonl but no oura-raw.jsonl — because the raw writer had only ever lived on a feature branch and was never merged to main. The bundle assembler already lists oura-raw.jsonl in its attach-if-present set, so on main it silently attaches nothing. This lands the missing writer.

• OuraRawDumpLine (WhoopStore, pure) + tests: byte-stable {schema,deviceId,utc,iso,hex} line, contiguous lowercase hex. 4 tests green.
• OuraRawDump (app): append-only writer, no dedup (a re-serve is still evidence the ring re-sent it; the offline reframer collapses by tag+ring-time).
• OuraLiveSource: rawDump property + tap on both history TLV feed points (secure and non-secure). Same feedsLive && !deviceId.isEmpty gate as the activity/motion dumps; the live-HR push is excluded so a night stays bounded.

Diagnostic only, never scored, never read back, safe to delete. Cross-platform: the decoded Oura sidecars have Android twins, but this raw capture is Apple-only in origin and is a diagnostic artifact (not stored data / analytics), so it is not parity-gated; the Android twin is a fast-follow and the only parity-sensitive surface is the pure OuraRawDumpLine encoder, which will be mirrored byte-for-byte.

Verified: WhoopStore OuraRawDumpLine 4/4 green; Strand macOS app BUILD SUCCEEDED (app-target Swift — no default CI covers it).

[ryanbr]

Makes sense, and the gap is real — the bundle assembler already looks for oura-raw.jsonl, so right now it attaches nothing. Thanks for building the macOS app too; nothing in CI covers that path.

One thing before merge: this file has no size cap. The activity and motion dumps both rotate at 25 MB (#676), and this one needs it more than either — it stores full hex of every notification and deliberately has no dedup, so it grows faster than both. A ~12 MB oura-raw.jsonl already hung the report review sheet once.

Can you lift the rotation from OuraMotionDump? It's about 15 lines and drops straight in. Happy to merge once that's there.

[pipiche38]

Makes sense, and the gap is real — the bundle assembler already looks for oura-raw.jsonl, so right now it attaches nothing. Thanks for building the macOS app too; nothing in CI covers that path.

One thing before merge: this file has no size cap. The activity and motion dumps both rotate at 25 MB (#676), and this one needs it more than either — it stores full hex of every notification and deliberately has no dedup, so it grows faster than both. A ~12 MB oura-raw.jsonl already hung the report review sheet once.

Can you lift the rotation from OuraMotionDump? It's about 15 lines and drops straight in. Happy to merge once that's there.

@ryanbr One more thing worth flagging alongside it: #702 (the master "Diagnostics capture" toggle) is arguably just as important here as rotation, for a different reason — rotation bounds the file after it's growing, #702 stops it from being written at all when the user hasn't opted into diagnostics. That matters more for this file than for activityDump/motionDump: it's full undecoded hex of every notification with no dedup, so it's the one that most benefits from not
existing by default.

Worth noting: #702 currently wraps activityDump + motionDump only — rawDump didn't exist on main when that PR was built, so it isn't wired into the gate yet. Once #702 merges, OuraRawDump's instantiation in OuraLiveSource.swift:785 will need the same diagnosticsCapture check, or this file stays always-on while its siblings are opt-in.

[ryanbr]

Rotation's in and it's the same logic as OuraMotionDump line for line, including the fresh-stat note about
the cached URL size. One commit, nothing else touched.

Worth knowing: the 20 MB figure I mentioned was wrong in the code, not in yours — ReportReviewGate said
20 while both dumps enforce 25. Fixed that in #939, so your 25 now agrees with the comment.

Approving. Still keen on the Android twin when you get to it.