fix(gemini): avoid duplicate structured schema prompt

[r266-tech]

Summary

• stops adding the structured-output JSON schema to Gemini system_instruction when native response_schema is already sent
• keeps response_schema on normal and cached structured calls
• adds a prompt-schema compatibility retry only for uncached structured calls that return invalid JSON, so cached retries keep their cached context

Fixes #1070.

Tests

• uv run --project hindsight-api-slim pytest hindsight-api-slim/tests/test_llm_extra_body.py -q
• uv run --project hindsight-api-slim ruff format --check hindsight-api-slim/hindsight_api/engine/providers/gemini_llm.py hindsight-api-slim/tests/test_llm_extra_body.py
• uv run --project hindsight-api-slim ruff check hindsight-api-slim/hindsight_api/engine/providers/gemini_llm.py
• git diff --check

Codex adversarial review approved after reworking the cached retry path so it does not drop cached context.

[koriyoshi2041]

I did a quick local check because the only red job is verify-generated-files. The failure appears unrelated to this Gemini change: that run dies while parsing hindsight-docs/src/data/integrations.json with the trailing-comma error that was fixed on main by 55f70e1d.

Targeted checks on this branch passed for me:

uv run --project hindsight-api-slim pytest hindsight-api-slim/tests/test_llm_extra_body.py -q
uv run --project hindsight-api-slim ruff format --check hindsight-api-slim/hindsight_api/engine/providers/gemini_llm.py hindsight-api-slim/tests/test_llm_extra_body.py
uv run --project hindsight-api-slim ruff check hindsight-api-slim/hindsight_api/engine/providers/gemini_llm.py
git diff --check

Result: 19 passed, ruff clean, no diff whitespace issues. A rebase/rerun should be enough to separate this PR from the stale docs failure.

[r266-tech]

@koriyoshi2041 confirmed — thanks for the targeted verification. The red verify-generated-files is the stale hindsight-docs/src/data/integrations.json trailing-comma parse error fixed on main by 55f70e1, not this change (which only touches gemini_llm.py + test_llm_extra_body.py). I've updated the branch onto current main so the check re-runs against the fixed base.

[nicoloboschi]

Approving. Verified the structured-output behavior with a Gemini LOCOMO run on conv-26 (this branch, isolated DB): 89.5% accuracy (136/152) and zero invalid-JSON / prompt-fallback events across ~70 ingestion fact-extractions + 152 recall/QA cycles. Native response_schema-only path is robust; removing the duplicate prompt-side schema does not degrade Gemini output.