feat(chat): route Gemma via LiteLLM, add Qwen 3.6, deadline + fallback

[Movm]

Summary

Regolo's gemma4-31b endpoint hangs upstream — every notebook chat with the default Gemma model spun forever because the AI SDK has no built-in first-token deadline. This PR fixes the immediate hang, adds resilience for similar future outages, and adds Qwen 3.6 27B as a new selectable model.

What changes for users

• Gemma 4 still appears as a model option but now routes through LiteLLM (Verdigado-hosted gemma4) instead of Regolo. No UI change; old gemma-regolo selections are migrated to gemma-litellm on next page load.
• Qwen 3.6 27B is a new selectable model. Reasoning streamer wires up automatically (already in REGOLO_REASONING_MODELS).
• If a model fails to produce a first content token within 20s, the server silently falls back to a sibling on the other provider. No user-visible UI for fallbacks — just a log line in the browser console (per request from product).

Chinese-only-when-selected firewall

This is the load-bearing constraint of this PR. Qwen entries in AVAILABLE_MODELS intentionally have NO fallback field. The firewall works both directions:

• Never auto-route INTO Qwen (would violate informed-consent boundary — users see the explicit "Chinesisches Modell – unterliegt staatlicher Zensur" warning before opting in)
• Never silently auto-route OUT of Qwen (the user picked Qwen for a reason; substituting hides the failure)

Encoded in code, not just docs: streamWithFallback early-returns when primary.config?.fallback is undefined. Reviewers can verify the firewall by reading AVAILABLE_MODELS — Qwen entries omit the field, comment block at `ModelConfig.fallback` explains why.

Pre-existing bug fixed

`getModel('litellm', modelId)` ignored the modelId argument and always used `LITELLM_DEFAULT_MODEL`. Latent footgun — masked today because both LiteLLM-mapped IDs (`litellm`, `gemma-litellm`) want the same physical model. Adding any second LiteLLM-served model would have silently routed to the wrong one. One-line fix included.

Implementation highlights

• New `litellmFetchWithThinkingDisabled` (sibling to `regoloFetchWithThinkingDisabled`) injects Ollama's top-level `think: false` so gemma streams content into `delta.content` instead of burning the entire token budget on `reasoning`.
• 20s first-token deadline guards against silent upstream hangs. Single shared deadline across initial probes (was accidentally giving 40s grace via per-call setTimeout).
• Reasoning streamer split into Phase-1 (race vs deadline until first text chunk) + Phase-2 (drain without race) — eliminates per-chunk Promise.race microtask hops after first content arrives.
• Native `AbortSignal.any()` (Node 20.3+) replaces a hand-rolled `composeAbortSignals` helper.
• `streamAndAccumulate` / `streamAndAccumulateWithReasoning` keep their old null-on-failure shape for existing chat router callers via a shared `wrapWithCompatCatch` factory; they get the deadline + empty-completion safety nets for free.

Out of scope (flagged for follow-up)

• Vision pipeline also uses broken Regolo gemma (`providers.ts:20` — `VISION_MODEL = { provider: 'regolo', model: 'gemma4-31b' }`). Same outage hits image analysis; same fallback policy could apply. Separate PR.
• Worker pool / monitor / voice paths also call `getModel()` but are non-streaming, so failure shape differs. Bigger refactor.

Test plan

• Manual: Berlin notebook chat with Gemma 4 (default) returns an answer. Pre-PR: spins forever. Post-PR: gemma streams via LiteLLM (~10s reasoning preamble swallowed by AI SDK, then content streams).
• Manual: Pick Qwen 3.6 27B in the model picker — answer streams with visible reasoning chunks via the existing reasoning UI.
• Manual: Pick Qwen 120B and ask a question — works as before; if Regolo Qwen happens to be down, error message says "Antwort konnte nicht generiert werden" (no silent fallback to non-Chinese model).
• Manual: With browser DevTools open, force Regolo to fail (block `api.regolo.ai` in Network panel) and pick GPT-OSS — server falls back to LiteLLM gemma; no UI banner; `[Notebook] Model fallback: gpt-oss-regolo → gemma-litellm (first_token_timeout)` appears in console.
• Persistence: load app with stale localStorage `selectedModel: 'gemma-regolo'` — chatStore migration v6 upgrades to `gemma-litellm` automatically.
• Regression: existing chat routes (`chatGraphContractRouter`, `chatGraphController`, `searchGraphController`) still work — they use the backward-compatible `streamAndAccumulate` wrapper; behavior unchanged plus the new 20s deadline as bonus safety.
• `pnpm --filter @gruenerator/api typecheck`
• `pnpm --filter @gruenerator/chat typecheck`
• `pnpm --filter @gruenerator/api test` (specifically `responseStreaming` if any tests exist)

Notes

• A duplicate of this commit exists on `feat/notebook-startpage-tabs-and-stats` (PR refactor(notebook): startpage tab/label cleanup + ts-rest auth status #693) because it was committed there first before being moved to its own branch. When this PR merges, the duplicate becomes a no-op for refactor(notebook): startpage tab/label cleanup + ts-rest auth status #693's rebase.
• This change introduces no new tests — the streaming layer is hard to unit-test against a real LLM. Worth adding mocked vitests for `streamWithFallback` (4 cases: success, primary-timeout-fallback-success, both-fail, no-fallback-configured) as a follow-up.