release/v1.28.31 — heart-rate history keeps its daily shape

[MBombeck]

Dense intra-day retention (PULSE / HRV / SpO2) now folds to hourly means instead of a single daily mean — the day's shape survives permanently (up to 24 points/day ≈ 9k rows/metric/year), while the last 14 days keep every raw sample as before.

Fold changes

• Hourly key stats:<HK>:<YYYY-MM-DD>T<HH>, local HH:30 anchor; stays under stats: so the idempotency predicate keeps holding; never collides with the daily shape or iOS ISO-instant wire ids.
• Same adopt-in-place slot logic against both unique indexes (index-A externalId first, index-B anchor instant second), P2002 caught outside the transaction with one retry.
• Pre-fold DAY-rollup recompute extended to HRV (the real constraint is mean fidelity — a post-fold recompute over hourly rows would store an unweighted mean-of-means); post-fold recompute removed tier-wide.
• Derived-resting minting for proxy users unchanged (still from raw).

One-shot history rebuild (pg-boss, boot-discovered, once per install)

• Per qualifying day (live daily-grain row older than the window + surviving tombstoned raws): mint hourly rows and retire the daily row in the same transaction — no instant with both grains live, so no double-count.
• The retired daily row is the durable marker; discovery converges to zero. Days with pruned tombstones keep their daily mean.
• PULSE/SpO2 DAY buckets untouched (true pre-fold stats); HRV DAY recomputed from hourly rows; W/M/Y whole-span enqueues.
• Discovery SQL parameter-bound, constant regex only, pinned by integration tests against real Postgres.

Consumers audited: stress engine, readiness, coincident-deviation, day-detail chart paths — none assume one row per out-of-window day.

Gate: typecheck, lint, 13537 unit tests, dense-tier integration suites (real Postgres, both P2002 vectors + atomic retirement + convergence), prettier, knip, build, openapi:check — all green.