Add records, Elijah scores, per-session notes, and PWA update prompt

App features:
- Personal-best records per metric: manually settable in Settings and
  auto-updated when a session beats them; shown in the log modal and a
  new dashboard "Personal records" card.
- Juggling now counts by 1 instead of 5.
- 1-on-1 with Elijah gains Technical Skill and Effort scores (out of 10)
  as manual inputs, plus an optional per-session note.
- Service worker now uses a controlled update flow: an in-app
  "new version ready" banner activates the update on tap and reloads.

Data model:
- category_metrics gains track_record + record; entries gains note.
- Idempotent migrations bring existing databases up to date (juggling
  step/record, Elijah score metrics) alongside the updated seed.

StartOS package:
- Bump to 0.1.2:0 with release notes.
- Build x86_64 only (drop aarch64) per deployment target.

s9pk/TODO.md

@@ -1 +1,11 @@
 # TODO
+
+## Known follow-ups
+
+- **In-app password change vs. StartOS action.** When deployed on StartOS, `PG_PASSWORD` (from `store.json`) is authoritative and re-applied on every restart. A password changed through the app's own Settings screen will be reverted on the next restart. Either hide the in-app password field under StartOS, or have the app write changes back to `store.json` so StartOS picks them up. For now, documented to use the **Set Login Password** action.
+- **Vendored app is git-ignored.** `./app` is regenerated by `make prep` and excluded from git, so app source changes do not bump the package `gitHash`. Fine for single-user deployment; revisit if publishing.
+- **riscv64 not built.** Manifest declares only x86_64 + aarch64 (Node + better-sqlite3 prebuilds). Add riscv64 only if a target host needs it.
+
+## Phase 3 (deferred — not started)
+
+- DGX Spark (Qwen3.6 35B, OpenAI-compatible endpoint) AI coach: login-time training suggestions and per-category drill ideas.