[pipiche38]

Two export-bundler bugs, one file

1. Three Oura sidecars never made it into the bundle. TestBundleAssembler.normalizedOuraEntryName()
hard-coded the Oura diagnostics kinds it recognises to the original three (raw/ibihr/activity).
The cva-ppg, motion and real-steps dump writers shipped later and have been writing real files
on-device the whole time — but the bundler didn't know their kind, so it silently dropped all three
from every Test Centre export. Confirmed against two real exports where the strap log shows the writers
firing and the files never appear in the zip.

Fix: collapse the four places that independently listed sidecar kinds down to one source of truth
(ouraSidecarKinds), so a future dump writer needs exactly one line added to be picked up by
normalizedOuraEntryName / ouraDiagnosticEntries / trimmableNames / ouraSidecarNames together,
instead of silently missing export.

2. Oversized JSONL sidecars were trimmed to a raw byte count. capEntries keeps the most-recent tail
of an oversized file by byte count — fine for a freeform text log, but a byte-count cut on a
newline-delimited JSONL file can land mid-record. Confirmed in production: a real overnight export
(2026-07-30/31) shipped oura-cva-ppg.jsonl, oura-real-steps.jsonl and oura-motion.jsonl each with a
corrupted first line, the first time their combined size crossed the 20 MB cap.

Fix: trimToLineBoundary — after the proportional byte-count trim, drop bytes up to and including
the first newline, so every kept line is a complete record. Every trimmable name is JSONL, so this is
safe across the board; a single very-long line with no newline is returned unchanged, never worse than
before.

Verification

• StrandTests 1753/0 (1 pre-existing skip) on the rebased tip — app-target Swift, so verified via
  xcodebuild … test, not just swift test. TestBundleAssemblerTests 14/14, including the new
  testCapSnapsTrimmedTailToLineBoundary (every kept line parses as valid JSON after a forced mid-file
  trim) and a normalisation case for each newly-recognised sidecar kind.
• Apple-only, and deliberately so. Android does have OuraMotionDump/OuraRealStepsDump writers, but
  its TestBundleAssembler.kt exports no Oura sidecars at all today (only raw-capture.jsonl) — so
  the Android situation is a wider, pre-existing gap ("Oura sidecars are never exported"), not a twin of
  this bug ("three of six sidecars are dropped"). Worth its own issue; not smuggled into this PR.

[ryanbr]

This conflicts with main now, and the reason is friendly rather than a problem with it.

#2154 landed, which rewrote the same export path: capEntries shares the cap by water-filling instead of proportionally, OuraRawDump is scoped to one capture session, generations are kept as a ring, and OuraSidecarGenerations.mergePlan merges them oldest to newest under the ceiling. Your own #2154 description named this issue as the wider, pre-existing gap, so the two were always adjacent; the larger one simply went first and moved the ground.

Worth rebasing and then deciding what is left rather than replaying it. Some of what this addresses may already be covered by the merge, and the part that is not will read differently against the new shape.

The seven other Oura PRs merged clean, so main now also carries the 5 minute fetch cadence, the freshly-offloaded scoring, the onset-rekey guard, the SpO2 column change, the 10 second anchor gap and both documentation corrections. That is a lot of adjacent movement in one evening, which is another reason to rebase before re-reading this one.

No review yet on the substance: I would rather read it against the tree it will actually live in.

[pipiche38]

Rebased onto main@2a4c157 to see what was left: nothing. Both commits come out empty — #2154 was
built on top of this branch and carried the whole of it (ouraSidecarKinds, trimToLineBoundary,
the line-boundary test and the per-kind normalisation cases are all on main). The only conflicts
were the two doc-comment lines #2154 rewrote when it replaced the proportional trim with the
fair-share one.

Closing this as superseded by #2154 rather than re-landing an empty diff.