fix: stabilize local model review flow

[ftchvs]

Summary

• Stabilizes local model review runs so the UI recovers from slow or stuck Ollama calls instead of leaving the form disabled.
• Exposes installed local review models as real dropdown options and adds an All platform mode for broader policy checks.
• Bounds local API model generation and disables Ollama thinking output for classifier calls so installed review models can return parseable AdLint JSON.

Changes

• adlint/static/app.js adds timeout-controlled /analyze fetches, model option population, and review-model filtering.
• adlint/static/index.html changes model selection to a dropdown and adds the All platform option.
• adlint/policy.py and adlint/rules/engine.py make platform=all include platform-scoped policies and LinkedIn derived rules.
• adlint/classifiers/ollama.py sends think: false and accepts fenced JSON responses from local models.
• Makefile starts the local API with bounded Ollama timeout and token generation settings.
• Tests cover UI assets, model controls, all-platform filtering, timeout recovery hooks, deterministic Ollama payloads, and fenced JSON parsing.

Test plan

• make test
• Live local-model matrix with ADLINT_OLLAMA_TIMEOUT=180 ADLINT_OLLAMA_NUM_PREDICT=1024: gpt-oss-safeguard:20b, gpt-oss:20b, qwen3-coder:30b, qwen3.5:35b-a3b, and gemma4:26b all returned model.status: ok
• Browser smoke at http://127.0.0.1:8000/ui/ with gpt-oss-safeguard:20b returning model.status: ok
• Browser smoke with gemma4:26b returning model.status: ok

Notes

• Embedding-only Ollama models remain filtered out of the UI because they are not review/chat models. Future model additions should still be validated against the strict adlint.model_review.v1 schema.

[ftchvs]

@claude please review