test: promote scratch pins — host-owned alarm, runSyncedWrite capture/broadcast seams

[grrowl]

Promotes two scratch verification files into the permanent suite by topic (per the signed-off triage table). The scratch files themselves are not added — this is a by-topic promotion, not a file move. Zero product code changes; tests only.

Row → landed test mapping

Pin	Verdict	Where it landed
alarm: no alarm armed through a full maybeCompact cycle	keep	host-matrix.test.ts → "a full maybeCompact housekeeping cycle arms no alarm"
alarm: host-owned alarm survives housekeeping	keep	host-matrix.test.ts → "a host-owned alarm survives a maybeCompact housekeeping cycle"
SW-A: callback arg === ctx.storage.sql	retire	folded as a one-line expect(sql).toBe(state.storage.sql) inside SW-B
SW-B: Drizzle-style direct ctx.storage.sql.exec broadcasts	keep	server-write.test.ts → "a Drizzle-style direct ctx.storage.sql.exec inside runSyncedWrite still broadcasts…"
SW-C: write outside runSyncedWrite captured but not broadcast until next drain	keep	server-write.test.ts → "a write OUTSIDE runSyncedWrite is trigger-captured but NOT broadcast until the next drain"
SW-D: multi-statement write → one wire batch, single uptodate	keep (trimmed)	server-write.test.ts → "a multi-statement runSyncedWrite reaches the client as one batch…"
SW-E: 30 rapid updates → 1 delta	retire	duplicate of coalesce.test.ts:39; its final-value assertion (coalesce.test.ts:69) already exists, so no strengthening was needed

Why these pin their ADRs

• Alarm tests (ADR-0015, host-owned alarm slot). The cohosting proof's third leg: tddc defines no alarm() — compaction rides ctx.waitUntil (#maybeCompact), so the DO's single alarm slot stays the host's. Both tests drive a real drain → #maybeCompact → waitUntil cycle over MaintTestDO (compactionEvery=3) and poll for the change-log collapse before asserting the alarm slot. The host-owned-alarm test is the stronger guard: the DO has a single alarm slot, so any tddc use of it would clobber the host's value — its exact survival proves the slot was never touched. (That the housekeeping actually ran is pinned in depth by maybe-compact.test.ts.)
• Server-write tests (ADR-0006, capture-vs-broadcast). SW-C asserts the "dark half" the suite only stated in prose — a raw sql.exec fires the CDC trigger (row lands in _sync_changes) but is not broadcast until a drain. SW-B proves the sanctioned path works even when a Drizzle driver writes through ctx.storage.sql directly. SW-D pins the wire framing (one batch, single uptodate, shared seq, one delta per key) that the collection abstraction hides.

Idiom / integration notes

• Rewritten to destination-suite idiom: shared helpers, runInDurableObject, unique-DO-name isolation. SW-B/SW-C reuse the file's existing collection-based helpers; SW-D adds a minimal codec-level socket (the only way to observe the wire boundary).
• SW-D trimmed per verdict (dropped redundant scratch assertions; kept the load-bearing batch invariants).

Adversarial review

Ran the diff past codex (gpt-5.6-sol) per repo practice; addressed its findings: hardened the batch test with an exact key-set assertion (one delta per key), reordered subscribeRaw to attach its listener before sending, widened the negative-broadcast window to several coalescer ticks, and corrected the alarm block's comment to claim only the observable contract (not the internal mechanism, which maybe-compact.test.ts owns).

Validation

• npm test: 269 passed (was 264 → +5), 53 files green.
• npm run typecheck: clean.

🤖 Generated with Claude Code