Make entities required with an empty-list default so literal models stop dropping them

[dimonnld]

Fixes #2749.

Problem

The four extraction models (ExtractedFact, ExtractedFactVerbose, ExtractedFactNoCausal, VerbatimExtractedFact) declared entities as an optional field that was absent from json_schema_extra["required"]. Under a strict JSON schema a literal model resolves the conflict — the field is optional and typed as objects, while the few-shot examples taught bare strings — by omitting it entirely: zero entities, no error. It bites precisely the most schema-conformant models and reproduces with strict-schema both on and off.

The retain parser is lenient (it coerces string entities into Entity(text=...)), so the string-array shape itself survives — the load-bearing defect is the omitted optional field, not the string shape.

Fix

• entities: list[Entity] = Field(default_factory=list, ...) on all four models, and "entities" added to each model's required array. An empty list stays valid, so lenient parsing never hard-fails, but the field is no longer silently omitted by literal models.
• Prompts teach the object form. The concise few-shot examples now use [{"text": ...}] objects, and an explicit "array of objects with a text field, never bare strings" instruction is added to the concise/verbatim (base) prompt and the verbose prompt — so non-strict / no-schema paths stop emitting the ambiguous shape.

Tests

test_fact_extraction_entities_schema.py — deterministic, no LLM:

• each of the four models' generated JSON schema lists entities as required, typed as an array of objects with a text field (the CI regression that would have caught this);
• omitting entities still parses, defaulting to [] (lenient path intact);
• every extraction prompt (concise / verbose / verbatim / custom) teaches the object form and shows no bare-string entity examples.

Verified locally: the schema/prompt suite is green, all prompt modes assemble, and the existing test_fact_extraction_fact_type_prompt.py still passes. Note the base prompt is .format()-ed twice (module assembly, then per-request), so the object examples in the few-shot are escaped accordingly and the standalone format instruction is brace-free.

[dimonnld]

Closing as superseded. #2749 was already fixed in main by #2830, which took the list[str] (flat strings) approach — main now has entities: list[str] = Field(default_factory=list). This PR used list[Entity] (required + empty-list default), which conflicts with that design choice, and its regression test duplicates the one #2830 added (test_fact_extraction_entities_schema.py). Nothing left to land here.