fix(gemini): grammar-enforce batch structured output regardless of strict flag

[benfrank241]

Problem

With LLM_PROVIDER=gemini and HINDSIGHT_API_RETAIN_BATCH_ENABLED=true, the Gemini batch request builder (_openai_body_to_gemini_request in gemini_llm.py) only set the native grammar schema (responseJsonSchema) when json_schema.strict was truthy. But HINDSIGHT_API_LLM_STRICT_SCHEMA defaults to False.

This created a batch-vs-interactive asymmetry: the interactive Gemini path always grammar-enforces structured output via its native response_schema regardless of strict (as its docstring states). strict is an OpenAI concept and is meaningless to Gemini.

As a result, at default config batch requests received only responseMimeType: application/json plus a textual schema hint in the system instruction — no native grammar enforcement. Gemini then intermittently emitted malformed JSON, and since a chunk's facts are parsed as a unit, all facts in the affected chunk were lost.

Fix

Set generationConfig["responseJsonSchema"] = schema whenever a schema is present, regardless of strict, so the batch path mirrors the interactive path. The always-set responseMimeType and the textual schema hint are unchanged. A comment references #2699.

Test

Updated the pure translation unit test in tests/test_gemini_batch.py (test_body_translation_grammar_enforces_schema_without_strict) to build an OpenAI-style body with strict=False and assert the resulting Gemini generationConfig contains responseJsonSchema equal to the schema. This replaces the prior test that asserted the buggy omit-when-not-strict behavior. No DB/LLM — runs fast.

27 passed in 7.26s

Fixes #2699