feat(models): switch default get() model from Gemini 2.5 to Gemini 3.5 Flash

[philipph-askui]

Summary

Gemini 2.5 (Flash, Flash-Lite, Pro) is being deprecated from the AskUI API. This PR swaps every 2.5 reference over to gemini-3.5-flash.

• Default AskUiGeminiGetModel, AskUIImageQAProvider and GoogleImageQAProvider to gemini-3.5-flash. The first two are what ComputerAgent().get() uses over the AskUI API.
• Add ModelName.GEMINI__3_5__FLASH and ASKUI__GEMINI__3_5__FLASH. The four 2.5 constants stay with a deprecation docstring so existing imports keep working.
• Update docstring examples, docs/04_using_models.md, docs/05_bring_your_own_model_provider.md and examples/model_providers.py. gemini-2.5-pro is dropped from the docs since the proxy does not serve a 3.5 Pro (returns 404).
• E2E get tests use the 3.5 constant; the six gemini_pro parametrizations are removed. VLM routing unit tests use 3.5 Flash as their Gemini id.
• Flash-Lite had no references in the repo.

Verification

• pdm run qa:fix clean (mypy, ruff format, ruff lint).
• tests/unit/model_providers: 81 passed.
• tests/e2e/agent/test_get.py against the live AskUI proxy: 26 of 28 selected tests pass, including PDF, XLSX, DOCX and response-schema cases on 3.5 Flash.
• The two failures (test_get_with_pdf_too_large_with_default_model, test_get_with_recursive_response_schema[askui]) are pre-existing and reproduce identically on main. They do not depend on the model id.

Known issue: latency (out of scope for this PR)

Schema-constrained get() calls on gemini-3.5-flash are bimodal upstream at Google: roughly half complete in ~3s, the rest stall for 20–50s. This reproduces through both the Vertex and OpenAI-compatible AskUI proxies, is not caused by retries or thinking, and is independent of schema shape. Plain JSON mode (no response_schema) is a consistent 2–4s. Measured latencies on a trivial prompt, 4 runs each:

Model	Mode	Median	Max
gemini-2.5-flash	response_schema	0.9s	1.0s
gemini-3.5-flash	response_schema	21.2s	48.7s
gemini-3.5-flash	no schema	3.3s	4.2s

E2E get tests now take 14–32s each against a 60s per-test timeout, so they may flake in CI until this is addressed (e.g. by switching the get model to plain JSON mode with client-side Pydantic validation).

🤖 Generated with Claude Code