fix(ai): preserve verified context windows across providers

[bloosqr]

Deep Research rejected prompts against models whose real context windows were missing from /models, or disappeared from the budget cache after five minutes. Preserve the last successful catalogue for the session and replace its snapshot on refresh, including smaller limits, removals and missing metadata. Isolate custom endpoints, including a configuration change while a catalogue request is pending. If custom discovery fails, retain its known windows while still offering manually entered model IDs.

Extend the original DeepSeek/Claude fix to Gemini (inputTokenLimit), Copilot (SDK catalogue refreshes), Cerebras (optional public metadata for authenticated models only), OpenAI, Xiaomi MiMo and OpenCode Go. Documented fallbacks match exact provider/model IDs; Go uses its own gateway limits, including smaller input ceilings. Unknown Claude IDs and Codex subscription models retain the conservative unknown-model cap. Local runtime allocations and advertised route limits take precedence. The Server catalogue uses the same generated shared contracts.

Validation:

• The new integration regression fails against the original PR when a 600,000-token advertised window drops to 200,000 after five minutes, and passes with this change.
• A second regression reproduces a custom discovery failure that incorrectly replaced an 8,192-token window with the 32,768-token unknown-model cap; it passes after preserving the successful snapshot.
• Production catalogue parsing, budget resolution and both completion guards run under Electron. HTTP SDK traffic is confined to loopback fixtures; Copilot is stubbed at its runtime boundary. Tests cover refreshed reductions, failures, invalid metadata, endpoint isolation, pre-dispatch overflow rejection and Copilot's internal catalogue refresh.
• AI/provider/context/subscription/retrieval regression suites passed, as did typecheck and lint. Server catalogue tests are included in root CI discovery; the generated module is checked against its TypeScript source.
• Model limits were checked against provider documentation and public catalogues on 2026-09-30; all 33 Go entries were compared with the current provider-scoped registry. No paid inference requests were made.