feat(book)!: half-hour vocabulary grid

[wavekat-eason]

Draft on purpose — do not merge yet. Merging releases @wavekat/flow-schema@0.0.11, and adopting that release before the daemon fleet is updated takes booking flows off the air. The gate is stated below and in the platform's docs/34 §2.

What

BOOK_GRANULARITY_MINS 15 → 30, in the TS package and its Rust twin, so appointment starts land on :00 and :30 only.

	before	after
universe per voice	111	63
Mon–Fri 09:00–17:00, 30-min appointments	44	29

Why it is staged, and why this half is second

required_assets folds vocabulary_refs(node) into the set a daemon must have on disk before it will arm a flow — and it computes that set from the constant compiled into its own build, not from anything the version carries.

So a daemon still on the quarter hour, handed a version published after this change, asks for bktime_0915, does not find it, and refuses to arm the flow. Not the booking step — the whole flow, silently.

The safe order is therefore:

1. Platforms narrow what they offer first (wavekat-platform #243, already open), leaving the frozen vocabulary a strict superset of what every daemon asks for and every caller hears. Callers get half-hour slots immediately, and no device is affected.
2. Fleets update to a build carrying this constant.
3. Then this merges, and the platform bumps its dependency.

Step 3 is the only one that can break anything, which is why this PR is a draft rather than a queue item.

The cross-language pin

Second commit, and worth reading independently of the grid. The vocabulary is computed twice — TypeScript and Rust — from two copies of one constant, and nothing compared them. Both languages' own tests pass happily while the two disagree, and the symptom of disagreement is not a red build; it's a phone line going quiet on the devices that updated late.

The corpus now carries an optional requiredAssets expectation and both runners assert exact-set equality against it. Exact, unlike the errors expectation beside it, because this set is the contract rather than a description of one. The Rust side also asserts that at least one case pinned something, so a corpus that quietly stopped pinning can't pass by doing nothing.

Verified by reverting the Rust constant to 15 on its own — the pin fails with required assets differ for v2/valid/clinic, which is exactly the drift it exists to catch.

Tests

cargo test --workspace 60 passing, clippy and cargo fmt --check clean; pnpm -r test 138 passing. Each affected grid case was re-derived rather than renumbered.

🤖 Generated with Claude Code

https://claude.ai/code/session_018stjkZJ4kxDViq1ENnVDmH