Add loadAllResult to surface corrupt records in collectionStore (#2728)

[atomantic]

Summary

collectionStore.loadAll() filtered out null records (a record whose index.json is missing, unreadable, unparseable, or rejected by sanitizeRecord), so every createCollectionStore consumer reasoned over a partial set with no way to tell "N records" from "N+K records, K corrupt." A silent undercount was indistinguishable from a complete read.

This adds a sibling loadAllResult() -> { records, failedIds } to the store, keeping the "which ids failed to load" signal. loadAll() becomes a back-compat flatten wrapper (loadAll = (await loadAllResult()).records), so no existing consumer changes. This mirrors the existing listOutcomesResult precedent and the CLAUDE.md "sentinel + validate — distinguish absent vs empty" convention.

The first consumer is tightened: layeredIntelligenceOutcomes.listOutcomesResult now reads via loadAllResult and reports read: false + partial: true + failedIds when any outcome record fails to load, instead of read: true with a silently-short list. Both existing consumers already treat read: false as unavailable:

• routes/apps/taskTypes.js — renders the "couldn't load" dashboard state instead of a lie ("nothing filed yet").
• autonomousJobs/layeredIntelligenceHooks.js — reports LI's merge rate as unavailable to the reasoner rather than feeding it a fabricated short history.

So a single corrupt outcome record now taints the whole read (we can't know which app's history it belonged to) rather than silently undercounting.

The other loadAll consumers (universeBuilder, pipeline issues/series, writersRoom, creativeCommissions, catalog user types, storyBuilder) are left on the loadAll back-compat wrapper unchanged.

Test plan

• server/lib/collectionStore.test.js — new loadAllResult suite: clean read (empty failedIds), a corrupt record surfaced in failedIds while readable ones still return, sanitizer-rejected records reported in failedIds, and loadAll remaining a faithful flatten of loadAllResult().records.
• server/services/layeredIntelligenceOutcomes.test.js — new partial-read test (one clean record + one corrupt record dir) asserting read: false / partial: true / failedIds and the surviving outcome; existing total-failure/non-array mocks retargeted from loadAll to loadAllResult.
• cd server && npx vitest run collectionStore layeredIntelligenceOutcomes — 97 passing. Lib barrel/README invariant test passing.

Closes #2728