feat(models): add native Google GenAI generateContent dialect

[jack-scitix-ai]

Type

• feature — new benchmark, task, or capability

Summary

• Add an active google_genai dialect for the Gemini Developer API's native generateContent protocol, with audited request lowering and strict streaming/non-streaming response lifting.
• Support explicit v1 and v1beta wire contracts, including version-specific operations, function schemas, final-request validation, and distinct runtime provenance.
• Support text and image input, structured output, thinking controls, logprobs, request seeds, function tools, Google Search, URL Context, and opaque continuation.
• Enforce model-family constraints and reject unsupported combinations before I/O, including Gemini 2.5 mixed tools and Gemini 3 thinking, sampling, candidate-count, call-ID, and signature requirements.
• Preserve provider-issued tool history and signatures across continuation replay, and verify the final PreparedRequest, including an exact audit of requested hosted tools.
• Recognize both MAX_TOKENS and CONTINUATION as truncation while preserving partial output and provider-verbatim finish reasons.
• Keep model-dependent capabilities gated on model-profile outcomes; Interactions API and Vertex AI / Agent Platform authentication and resource paths remain outside this dialect.

Related Issues

Refs #25

Test Plan

Automated

• Lint/format clean (pdm run ruff check && pdm run ruff format --check)
• Type check clean (pdm run ty check sieval/)
• Unit and integration tests pass (pdm run python -m pytest tests/unit tests/integration -q): 8420 passed, 84 skipped

Checklist

Required (all PRs)

• PR title follows conventional format (type(scope): description)
• No internal paths, credentials, or personal info in committed files
• AI-generated code has AI-Generated Code - <model> (<provider>) in module docstring
• No new upper-layer dependencies added to core/
• Deleted code verified — no remaining call sites depend on it

[ethan-scitix]

Request changes. The design is solid and the tests are strong (unit + integration pass; ruff and ty are clean). Two accounting bugs put wrong numbers in reports without raising, and coverage is below the core/ gate.

Blocking

1. Bug: wrong usage semantics (google_genai.py:2039). candidatesTokenCount excludes thoughts (totalTokenCount = prompt + toolUsePrompt + thoughts + candidates). But UsageStats.reasoning_tokens must be a breakdown of output_tokens. With prompt=20, candidates=150, thoughts=3000, the profiler reports a 2000% reasoning share, and every thinking call trips the total-mismatch counter. Fix: output_tokens = candidates + thoughts and input_tokens = prompt + toolUsePrompt.
2. Bug: Google filter stops aren't counted as truncation (meta.py:24). The set includes content_filter, but not Google's equivalents: SAFETY, RECITATION, BLOCKLIST, PROHIBITED_CONTENT, SPII, IMAGE_SAFETY, … So a filtered Gemini rollout reads n_truncated = 0. Add them, or document why they're left out.
3. Coverage: google_genai.py is at 90%, below the 95% required for core/. Most uncovered lines are the fail-loud raise branches (e.g. :878–942, :1383–1447, :1743–1765).

Non-blocking

• Question: CONTINUATION is counted as truncation and continuationToken is silently dropped. That's fine for scoring, but please note it in infer.md and add a comment in the code.
• Nit: cast(Any, group) and int(cast(int, minimum)) in dialect.py aren't needed; ty passes without them. The AssertionError at dialect.py:~429 is unreachable.
• Nit: the body of commit 6f81a283 contains a literal \n\n.
• Follow-up: model-family limits are matched by regex on the model name, so gateway aliases skip them. Consider moving them into model profiles, and splitting the 3.8k-line module.