Restore CAIL generation and curate Jeopardy models

[milwrite]

Jeopardy sent full provider IDs to CAIL after its catalog switched to short aliases, causing generation failures. Resolve saved IDs against a curated, active catalog, and enforce the same six-model selection in the menu and inference route. Use a task-focused subset of CAIL Featured plus requested MiniMax M3 and compact Gemma 26B A4B. Retired selections migrate to the current default for future requests while preserving existing boards.

The model selection applies to boards and Final Jeopardy, including personal-key provider mapping. Preserve deadlines for board, connection and Final requests; give MiniMax short requests at least 2048 completion tokens without reducing full-board budgets. Fix clue validation that incorrectly treated common grammar as answer leakage or rejected specific clues beginning with “this city.” Phone form controls now use at least 16px text, and the configuration panel respects the dynamic viewport.

The curation and mobile changes were deployed as 9051bde, Worker version fcc0762f-86c4-4566-a43e-5e9c9f1e2ea4. The existing game-suite task subsequently deployed stream recovery on top of that baseline; the MiniMax budget commit 945fceb is included in coordinated release e5637ed, Worker 7f5c63e3-14e1-4cc3-86f3-6b33cc6a2aec, verified at 100%. That later release also selects Mistral Small 4 for new sessions and preserves saved choices. This PR targets the Cloudflare game-suite branch.

Validation: 16 tests, TypeScript and production build pass. Phone-sized production Chromium verified the six active model choices, MiniMax selection, 16px fields, and no horizontal overflow or page errors. Earlier signed-in production acceptance generated and saved a full board and Final clue. Additional non-streaming probes produced parseable 30-clue boards with DeepSeek V4.1, Gemma 26B A4B and MiniMax; those probes did not exercise the full UI prompt or validator, and later provider latency varied. The follow-on release 913d1a8 completed full signed-in app acceptance: 30 clues generated in 8.466 seconds, autosaved by 8.871 seconds, and an answered clue restored in a fresh tab with D1 revision/model readback. Worker 2c8f69af-f2a8-4045-a097-5da564779a64 is independently verified at 100%. These are measured-run timings, not a latency guarantee; the full receipt is in PR #7. Model-generated factual accuracy was not independently audited.