fix(visualize): unblock Gemini 2.5+ and harden Visualize pipeline

[skinred78]

Fixes #489.

Summary

Six bugs that combined to break the Visualize capability on Gemini 2.5 Flash (and similar thinking-by-default models). Each is independently useful, but the user-visible symptom — Visualize output randomly truncated at ~370 chars — needs all of them.

1. Gemini 2.5/3.x reasoning-tokens default (root cause)

Gemini 2.5+ models burn most of max_tokens on internal "thinking" tokens by default. With max_tokens=4096, ~3900 tokens went to reasoning and ~160 came out as visible content, causing finish_reason=length on every multi-step pipeline (Visualize codegen + review, Deep Solve writing, etc.). See #489 for the reproducer with curl directly against Gemini's OpenAI-compat endpoint.

Default reasoning_effort="none" for Gemini 2.5/3.x when the caller doesn't specify, in all three execution paths:

• provider_core/openai_compat_provider.py:_build_kwargs (live path)
• executors.py:sdk_complete / sdk_stream (SDK path)
• cloud_provider.py:_openai_complete / _openai_stream (aiohttp fallback)

Callers that want thinking can still opt in via explicit reasoning_effort.

2. visualize capability had no agents.yaml entry

get_agent_params("visualize") silently fell through to the 4096-token default because there was no entry in DEFAULT_AGENTS_SETTINGS or loader.py:section_map. Added both, with a 16384-token budget appropriate for full HTML pages.

3. Review stage crashed hard on JSON parse failure

ReviewAgent.process does ReviewResult.model_validate(extract_json_object(response)). When the model returned prose instead of JSON (common with large SVGs that don't escape cleanly into a JSON string field), the parse raised and killed the entire Visualize turn — the user saw the streamed SVG draft and then a traceback. Wrapped pipeline.run_review() in try/except so a parse failure falls back to the unreviewed draft and the user still gets a rendered result.

4. Codegen output not trimmed to the root tag

Models often wrap SVG/HTML in prose ("Here you go: … Enjoy!") or emit a closing code fence on the same line as </html>, which extract_code_block's regex (requiring a leading \n before the fence) doesn't strip. Added defensive root-tag trimming for render_type=="svg" and render_type=="html" so the renderer always sees a clean payload.

Test plan

Verified end-to-end on Gemini 2.5 Flash via the CLI and via headless Playwright:

• Long-division HTML test prompt: 22 KB self-contained interactive page (was: 0 bytes / crash on dev).

• All step prompts read correctly (e.g. "How many whole times does 6 go into 7?" for the first digit).

• Walking the full 7852 ÷ 6 algorithm step by step terminates with Quotient: 1308, Remainder: 4.

• Wrong-answer retry preserves prior progress; 3 wrong attempts reveals the answer and auto-advances (graceful-fallback path exercised).

• Simple SVG (e.g. "draw 8 cookies") still renders unchanged — no regression on the happy path.

• The defensive root-tag trim is a no-op when codegen already returns a clean tag.

• Maintainer: please run pre-commit run --all-files and CI as a sanity check; my local environment doesn't have all the lint deps configured.