* feat(runtime): device-anonymous learnerKey + app RuntimeStore singleton (#869)
* feat(runtime): cascade stage deletion into the runtime store (#869)
deleteStageWithRelatedData now cascades into the runtime layer after its
Dexie transaction completes. The runtime data lives in a separate
IndexedDB database (maic-runtime), so it cannot join the transaction, and
the cascade goes through deleteStageRuntimeSafely — a helper that never
throws (warns with context instead): a broken or hung runtime DB must not
brick stage deletion in the main app DB. deleteStageRuntime is idempotent,
so a failed cascade can simply be retried.
Covered at the helper seam with a stub RuntimeStore (success + throwing);
the repo has no Dexie-in-node harness for database.ts and this change
does not invent one.
* fix(runtime): wire the live deletion path, serialize learner-key minting (#869)
Cross-review fixes on the C1 bootstrap:
- The runtime cascade was wired into deleteStageWithRelatedData, which has
zero callers; the UI classroom-deletion flow goes through deleteStageData
in stage-storage. Wire deleteStageRuntimeSafely into that live path too
(after its Dexie work, same isolation rationale), and cover the wiring by
running the real deleteStageData with its module deps mocked — the
repo's established pattern for database-touching code.
- getLearnerKey minted twice under concurrency. Same-bundle callers now
share one in-flight promise on the default path (failures are not
cached), and every path re-reads after writing and returns the PERSISTED
key, so a cross-tab race converges on the stored winner instead of
keeping an orphaned local mint.
- Guard crypto.randomUUID with the house fallback pattern.
- Honest docs: the cascade comment no longer claims a retry path exists
(orphaned rows are inert today; a startup sweep is deferred to Part C2),
and both lib/runtime modules note they are client-only.
* fix(runtime): cross-tab mint lock and bounded deletion cascade (#869)
- Read-after-write alone still let a tab keep an orphaned learner key
when its re-read landed before the other tab's write. Minting now runs
under the Web Locks API ('maic:learner-key') where available: grants
are mutually exclusive across tabs, the loser re-reads the winner's key
inside its grant, and an existing key is never overwritten (so the
per-tab memo stays safe). Without navigator.locks (older browsers,
non-window contexts) the memo + read-after-write behavior remains, with
the residual race named in a comment and accepted — it merely splits
one anonymous learner's local history.
- deleteStageRuntimeSafely awaited without a bound, so a hung runtime
IndexedDB could block the live deletion path — the exact failure the
helper exists to isolate. The cascade now races a 5s timeout: on
timeout it warns and resolves (orphaned rows stay inert), and the still
-pending cascade carries a swallow handler so a late rejection cannot
become an unhandled rejection.
* fix(runtime): probe for the runtime DB before cascading a stage deletion (#869)
deleteStageRuntimeSafely reached openDb() unconditionally, and opening
CREATES the maic-runtime database — so deleting a classroom on a device
that never wrote runtime data paid an open-or-create of a second
IndexedDB DB, and in degraded environments burned the full 5s bound for
zero cleanup value.
Probe first without creating: where indexedDB.databases() is available,
a missing maic-runtime entry returns immediately; where the probe API is
unavailable (older Firefox), fall through to the bounded cascade —
skipping there would strand real cleanup once Part C2 adds writers. The
probe shares the existing try/catch + timeout envelope, so a hanging
databases() cannot brick deletion either. The DB name is a module const
passed explicitly to BrowserRuntimeStore so probe and store can never
drift.