feat(query): default invalidations to the keys optimistic() declares

[zzal]

Part II Wave 2 · #1 — delete the boilerplate, make the footgun inexpressible

Every optimistic CRUD mutation ended with the same required boilerplate restating keys the engine already knew:

func optimistic(_ i: Input) -> [OptimisticEdit] {
    [.update(TodoList()) { /* … */ }]
}
func invalidations(input: Input, output: Todo) -> [Invalidation] { [.exact(TodoList())] }  // ← restated, on every mutation

…and forgetting the method was the exact trap a DEBUG diagnostic papered over: the cache keeps the optimistic guess and never reconciles with the server.

Change

Mutation.invalidations now defaults to the keys optimistic(input) declares — .exact per declared key, deduped in declaration order:

• Derived from declarations, not the rollback stack (the audit's own caveat): a .noValue-skipped edit — nothing cached yet — still gets its refetch, so the server value lands either way. Deriving from rollback would silently lose exactly those (issue bug: optimistic mutation before its query loads traps in DEBUG (.noValue should be a benign skip) #96's territory). Test-pinned.
• Dedup matters: duplicate invalidations of one key cancel-respawn its repair fetch. Test-pinned.
• Not source-breaking: existing overrides win; an explicit [] override is the deliberate opt-out.
• Diagnostic retired: the optimistic-without-invalidations DEBUG trap + its exit-test suite are deleted — the misuse is now inexpressible (the derived default can't be empty while optimistic edits exist). Making misuse inexpressible beats warning about it.
• New documented contract: optimistic(_:) should be a pure declaration (the default re-reads it on success to learn the keys); run-once effects — like allocating an optimistic temp id — belong inside the edit's transform closure, which runs only when the edit applies.

Dogfood

All 4 invalidations overrides in the repo deleted (TodoCRUD ×3, QueryDemo ×1). AddTodo's temp-id allocation moved into its transform closure per the purity rule. EmbeddedTemplates.swift regenerated.

TDD

RED reproduced the footgun end-to-end before implementing — fetches → 1 (no repair refetch) and the cache stuck on the optimistic guess [100, 7]. 5 new tests: derived-default shape (dedup + order), skipped-edits-still-contribute (the caveat), override-wins, empty-optimistic → empty, and end-to-end reconciliation via the derived refetch.

Verification

• Full host suite: 1383 tests, 277 suites green
• TodoCRUD + QueryDemo wasm builds clean
• Release CLI rebuilt, then TodoCRUD Playwright e2e 5/5 against the real Bun backend scaffolded from the new templates — "optimistic add reconciles to the server row" now exercises the derived invalidation in a real browser; forced-failure rollback unaffected

🤖 Generated with Claude Code

[github-actions]

📦 Bundle size

Artifact	Baseline	This PR	Δ
App.wasm	5.06 MB	5.15 MB	+1.90%
App.wasm (gzip)	1.80 MB	1.84 MB	+2.17%
JS runtime	54.5 KB	54.5 KB	+0.00%
JS runtime (gzip)	11.3 KB	11.3 KB	+0.00%
Total (gzip)	1.81 MB	1.85 MB	+2.16%

✅ Within budget (≤5% growth allowed).

_{Baseline: Swift 6.3, WASM SDK 6.3-RELEASE, measured 2026-06-18.}