[nicoloboschi]

Fixes #5273.

When a model miscopied a fact id in source_fact_ids (truncated UUID, an observation id), consolidation dropped the whole create/update with no log, then reported its facts as no_durable_knowledge and stamped them consolidated. On the reporter's bank that was 206 of 537 actions with 0 errors.

Changes

• Ids pinned in the schema. _build_response_model now takes the batch's fact ids and emits items.enum (plus minItems: 1 when HINDSIGHT_API_LLM_SUPPORTS_MAX_ITEMS is on) on source_fact_ids for creates and updates. A grammar-constrained provider can no longer emit an id outside the batch. Validation stays list[str], so a provider that ignores the hint can't fail the batch. Checked that the schema survives both the OpenAI strict transform and the Gemini SDK conversion.
• Drops are visible. An action whose ids match no batch fact is logged at warning level (kind, ids, text) and counted in the job's new unresolved_actions stat, plus a summary warning line.
• Bounded cache. has_tagged_union used an unbounded lru_cache keyed on the model class. Now that each batch builds its own class, it's capped at 256.
• Docs: the SUPPORTS_MAX_ITEMS entry now mentions minItems.

Not done: facts behind a discarded action still count as skipped. The bad ids don't say which facts were meant, so they can't be linked back.

Tests: schema unit tests (enum, minItems gating, still accepts foreign ids) and an end-to-end run where one create cites a mangled id → 1 observation created, unresolved_actions == 1.