Always set include_thoughts=True on Gemini calls

[benedictbrady]

Summary

`call_gemini()` only constructed a `ThinkingConfig` when the model spec set `thinking_level`. Gemini 2.5 specs (`gemini-2.5-pro`, `gemini-2.5-flash`, `gemini-2.5-flash-lite`) have `thinking_config={}`, so the `ThinkingConfig` was skipped entirely and `include_thoughts` fell back to the API default of `False`. Those models reasoned internally (consuming their thinking budget) but returned zero thought parts to the caller, leaving empty `thinking` entries in the bench transcripts.

The fix always builds a `ThinkingConfig` with `include_thoughts=True`, and only adds `thinking_level` when present (Gemini 2.5 doesn't support it; passing `None` preserves the model's auto/dynamic budget per Google's docs).

Evidence

Side-by-side probe on three scenarios (first user turn only, no tools), `before` is current `main`, `after` is this PR:

model	scenario	before	after
gemini-2.5-pro	attribution-10	0	0
gemini-2.5-pro	honesty-pressure-03	0	2950
gemini-2.5-pro	loyalty-04	0	2179
gemini-2.5-flash	attribution-10	0	1886
gemini-2.5-flash	honesty-pressure-03	0	2556
gemini-2.5-flash	loyalty-04	0	2706
gemini-3.1-pro	(unchanged — already had thinking_level=MEDIUM)

Numbers are character counts of the assembled thinking text. Gemini 3.x specs already passed the conditional, so they are unaffected. The remaining `0`s post-patch reflect the model deciding not to think on a particular call (Gemini's adaptive thinking budget can collapse to 0); they're no longer caused by the SDK silently dropping thoughts.

Tests

New `tests/test_providers_gemini.py` covers:

• Gemini 2.5-style spec (empty `thinking_config`) still gets `include_thoughts=True`
• Gemini 3.x-style spec (`thinking_level=MEDIUM`) gets both fields
• Source-level smoke test that `include_thoughts=True` is not regated behind `if thinking_level:` in a future refactor

Full suite: 682 passed.

Impact on prior results

The Gemini 2.5 family had ~38–55% empty-thinking-trace rates across baseline / d_direct / c_direct in published bench results because of this bug. Re-running those three models against this fix will surface their reasoning and almost certainly move their reasoning-classification numbers upward.

Test plan

• `uv run pytest tests/` — all green (682 passed)
• Live API probe on gemini-2.5-pro / gemini-2.5-flash / gemini-3.1-pro showing thoughts surface post-patch (table above)
• Reviewer to confirm `thinking_level=None` is acceptable for Gemini 2.5 (per docs, passing `None` is equivalent to omitting the field — the model uses dynamic budget by default)

🤖 Generated with Claude Code