v0.43.0 — uma chave para qualquer provedor, e as telas que faltavam

The release where a key for any provider LiteLLM reaches starts working, three things that existed
only in .env got a control, and two features that were built but unreachable finally have a
surface: attaching a file, and the registry the board dispatches to.

Added

• Any provider LiteLLM understands. LiteLLM is a hard dependency that reaches a hundred-odd
  vendors and Chimera recognised five. Someone holding a valid Groq or Mistral key was told "No
  provider key configured" and never reached LiteLLM at all — a refusal that, from outside, is
  indistinguishable from a broken product. The gate now accepts any <PROVIDER>_API_KEY.
  • Minus a denylist of our own credentials that are not sources of models. That list is mandatory
    rather than tidy: ELEVENLABS and STABILITY are real entries in LiteLLM's provider enum, so
    without it someone whose only credential is text-to-speech would pass the gate and doctor
    would name them as a provider of models.
  • A key that lived only in the .env was read and silently dropped, because Settings is
    extra="ignore" — it became neither an attribute nor an environment variable, and LiteLLM never
    saw it. Since chimera init writes a .env and the docs point people at it, opening the gate
    alone would have traded one error for another.
  • doctor flags a discovered name LiteLLM does not recognise, because a typo looks exactly like a
    provider. That check pays LiteLLM's import, which is why it lives in the diagnostic command and
    not in the gate.
• The first-run wizard picks a provider, and pins a model that works with it. Every cost preset
  is an OpenRouter slug, so an Anthropic key with the built-in default untouched left the tier
  ladder pointing at a vendor that key cannot reach — and the Test button then blamed the brand-new
  key for OpenRouter's 401. The suggestion is shown in an editable field rather than applied
  invisibly, because these slugs perish and a wrong one should be something you see before Save.
• The agent registry gets a screen, as Automation's third tab: list, create, edit, delete. The
  registry existed, was persisted and was dispatched against, and was reachable only from the CLI or
  a hand-written curl — so the board's lane box was a free text field guessing at a list the app
  never showed, and the first news that an id was wrong arrived as "0 worked", after the run.
  • The lane stays free text: filing work against an agent you have not created yet is reasonable,
    and the card waits in the backlog until it exists. The board now offers the registered ids as a
    datalist and says "no agent with this id yet" while you type.
• Three settings no interface could reach. The embed model, now directly under the
  semantic-memory switch that depends on it — recall degrades to lexical on any embedder failure,
  silently and by design, so a default the user's key could not serve made that toggle confirm a
  change it had not made. The Ollama base URL, which api_base does not cover because that one is
  sent on every call and redirects the cloud providers too. And key pools, which rotated round-robin
  with a per-key cooldown and had no control anywhere.
  • Pools are edited by operation, never by value: add takes one key, remove takes a position. A
    text field holding the CSV would have to display its value to be editable, the value is secret
    so it would display the mask, and one Save would write …abcd over a working rotation. There is
    no request shape here that carries a key back to the server, which makes that failure
    inexpressible rather than merely unlikely.

Fixed

• Attaching a document and dictating a message both answered 422 in the shipped app. Both send a
  FormData body through authHeaders(), which always sets Content-Type: application/json — and
  a Content-Type you write yourself is one the browser will not write for you, so the multipart
  boundary was missing. The bytes were still multipart; they just arrived labelled as something the
  server could not split, and FastAPI reported the file field as missing, which sends you looking
  at the form instead of at the header.
  • Cover was in the wrong place, which is why nothing noticed: every component test mocks
    @/lib/api wholesale, so the upload client was a stub and the real fetch never ran.
• The app lost your theme, workspace, project list and language on every launch. The shell
  started the backend with --port 0 and let the OS pick, so the window loaded a different
  http://127.0.0.1:<port> origin each time — and browsers partition localStorage by
  scheme+host+port. The port is remembered now and asked for again; asking is all it does, since
  the backend already falls back to a free port when the requested one is taken.
• Three of the five suggested model slugs were a generation behind, because they came from a
  catalogue that speaks OpenRouter's aliases rather than each vendor's native IDs. Checked against
  the vendors' own model lists.