fix(core): repair ~25 unit-test failures from the beta.25 cold-start caches

[mmcintosh]

Description

Three per-isolate caches / context assumptions introduced by the beta.25 cold-start work behave correctly in a live Workers isolate but break outside one — which is why 25 unit tests currently fail on a clean main checkout (npm test in packages/core). Each fix preserves the production optimization while making it correct across databases / runtimes. No behavior change on Workers.

Fixes #1016

Sequencing (PR 1 of 4)

This is the shared prerequisite for a 4-PR set — merge this first. The other three each re-ground onto this one; after it lands they touch disjoint files and can merge in any order.

1. PR 1 — fix(core): beta.25 test-infra (this PR) — repairs the ~25 pre-existing unit-test failures; the shared prerequisite.
2. PR 2 — feat: views plugin — re-grounds onto PR 1.
3. PR 3 — fix(bootstrap): first-boot FK race + single-flight — touches only bootstrap.ts; parallel after PR 1. (Its red CI before PR 1 lands is exactly the ~25 failures this PR fixes — none of them touch bootstrap.ts.)
4. PR 4 — feat: forms document-model port — re-grounds onto PR 1.

Changes

• document-scalar-schema.ts — the module-level PRAGMA column/index cache leaked across databases. In tests, the 2nd in-memory DB "already had" its q_* generated columns per the cache, the ALTERs were skipped, and every generated-column read failed with no such column: q_* (16 failures across 5 real-SQLite suites). Now a WeakMap keyed by the D1 binding: one bootstrap reuses one env.DB, so the intended per-isolate dedup is fully preserved, while a different database gets its own snapshot.
• plugin-middleware.ts — same class: isPluginActive's status cache was keyed by pluginId only, bleeding one mock DB's active into the next test's DB (7 failures). Now WeakMap-per-DB + a generation stamp for invalidatePluginStatusCache (a WeakMap can't be enumerated, so a bump lazily discards every DB's map — activate/deactivate is rare, the refill is one read per plugin).
• api.ts + catalog.ts — Hono's c.executionCtx getter throws when there's no ExecutionContext (vitest's app.request()), 500-ing the cached-collection read path (2 failures). Added a safeExecutionCtx() try/catch helper at all 9 call sites; scheduleKvWrite now tolerates an undefined ctx (the KV catalog warm is best-effort background work, never load-bearing).

Testing

Unit Tests

• Added/updated unit tests
• All unit tests passing

Verified on a branch off main containing ONLY these 4 files: tsc clean; the previously-failing test files now pass (6 files / 77 tests); full suite 1715 passed / 0 failed / 247 skipped (≈25 were failing on pure main). No bootstrap.ts / first-boot-sequence changes — this is test-harness isolation + a context guard, not a bootstrap concurrency change.

E2E Tests

• Added/updated E2E tests
• All E2E tests passing

Screenshots/Videos

Checklist

• Code follows project conventions
• Tests added/updated and passing
• Type checking passes
• No console errors or warnings
• Documentation updated (if needed)

[mmcintosh]

Note on the red CI. The only failing check is the E2E against preview job, and every failure is the same one: loginAsAdmin times out because the admin signs in but cannot reach /admin. This PR does no touch any auth or bootstrap code — it only fixes per-database test caches and an executionCtx guard. The cause is a separate, pre-existing first-boot bug: on a fresh preview database, bootstrap races a foreign-key dependency and silently leaves the RBAC roles and core plugins unseeded, so the admin gets a 403. That is fixed by the bootstrap PR (#1019, Fixes #1017), not here. This PR's own unit suite — the roughly 25 failures it fixes (#1016) — is green. The two are a pair: this PR fixes the bootstrap PR's unit tests, and the bootstrap PR fixes this PR's E2E preview. Recommend reviewing on the green unit suite; the E2E clears automatically once #1019 lands.