feat(providers): add Atlas Cloud provider

[ethanj]

Picks up #167, which had been open a month without a push. Rebased onto current main and finished against that review. @binyangzhu000-sudo is credited as co-author; the provider, the alias handling, and the original tests are their work.

What changed from the original PR

The default model. Atlas Cloud's catalogue marks tool support per model, and qwen/qwen3.5-flash carries no supported_features entry at all, while 35 of its 136 models advertise ["json_mode","structured_outputs","tools"]. compile extracts concepts through toolCall with tool_choice: "required", so that default would have failed on the very first extraction request rather than degrading — the provider would look installed and not work. The default is now qwen/qwen3.5-35b-a3b, the smallest catalogue model that does advertise tools, which keeps the cheap-and-fast intent of the original choice.

Docs. The three surfaces the repo requires for a user-facing provider: the LLMWIKI_PROVIDER enumeration plus a credentials section in environment-variables.mdx, a setup tab in providers.mdx, and the README provider table.

One list instead of two. SUPPORTED_PROVIDERS is now derived from SUPPORTED_PROVIDER_INPUTS rather than written out again. The original PR had grown the number of hand-maintained provider lists from two to three, and two of them drifting apart produces a guard that rejects what the factory can build. The stale module header in provider.ts, which had omitted Copilot and the Agent SDK since they were added, is fixed in the same pass.

Why a dedicated provider rather than the OpenAI-compatible escape hatch

All 136 catalogue ids are namespaced by publisher and bare gpt-4o is absent, so LLMWIKI_PROVIDER=openai with OPENAI_BASE_URL fails on its own default model.

The embeddings remedy, in all three providers

All three fail-closed providers told the user to change LLMWIKI_PROVIDER when embeddings are unavailable, which costs them the chat provider they picked. LLMWIKI_EMBEDDING_PROVIDER exists (#154) so they do not have to, and it is what the docs prescribe.

Atlas Cloud inherited that text verbatim from MiniMax, and a Copilot test asserted the wrong advice was correct — named embed() error message mentions switching to the openai provider, which is how it survived #154 shipping the variable that made it wrong. Fixing only the new file would have left the codebase giving two different answers and kept the test that made the wrong one look deliberate. All three now point at the embedding override, Copilot's names OPENAI_EMBEDDINGS_API_KEY rather than the main key, and the test asserts the correct remedy plus the absence of the old one.

Not included

Embeddings are not wired up and fail closed, matching MiniMax and Copilot. All three share an absent EMBEDDING_MODELS entry, which is better fixed once across all of them than for this provider alone.