feat(api): put any model provider behind one port

[Madeuss]

Groundwork for phase 2. ADR-0007
draws the boundary; ADR-0008
picks what sits behind it.

The interface

def complete(self, *, system: str, user: str, schema: type[Schema], max_tokens: int) -> Schema

A system prompt, a user prompt, a schema; an instance of it, or a failure. Two
failures kept apart: ModelUnavailable and ModelRefused — a refusal is a fact
about the request and belongs in the evals, not buried in generic plumbing.

Affordable because of decisions already made. Nothing in the deduction path
asks a model for anything: solver, verdict, scoring and contradiction detection
are code (RN-023, RN-032). An application whose business logic ran through tool
calling could not draw the line this tightly.

The provider

Magalu Prosa, which exposes an OpenAI-compatible API — so the adapter is not
a Prosa adapter. It is an OpenAI-compatible adapter with a base URL, and Groq,
Together, OpenRouter or a local vLLM are configuration away.

The grounds are in ADR-0008 and are not the ones ADR-0007 said it would wait
for: credits already held, infrastructure already on Magalu, a monthly ceiling
in R$ (the currency the eval plan is written in), and supporting a Brazilian
cloud. The ADR says that plainly rather than pretending the eval evidence
arrived early — and records why the port is what makes deciding early safe.

Structured output, which the docs do not cover

Prosa's documentation does not say whether the gateway enforces schemas. So the
adapter finds out at runtime and remembers:

1. response_format with a JSON schema — the server enforces the shape
2. JSON mode plus the schema in the prompt — valid JSON guaranteed, shape is the model's problem
3. plain request, JSON pulled out of whatever comes back

All three failing is a failure. A response that does not validate is discarded,
never repaired (RN-022).

What was removed, and why it should not have been there

The Anthropic adapter. It became dead code the moment the provider question was
raised — an adapter for a provider this project will not call, carrying an SDK
dependency for calls that will never happen. It survived because moving working
code behind the new port was easier than deleting it, which is not a reason.
git log keeps it.

The cost, stated plainly

Prosa's catalog is open weights — Google, Meta, NVIDIA, Qwen — and those models
are weakest exactly where phase 3 is hardest: injection resistance ≥ 95% and
persona consistency, in Portuguese, which ADR-0005 already flagged as the harder
language.

What does not degrade: the canary gate stays 0%. The model is never given a
secret fact, and the filter is code (RN-010, RN-012). A weaker model costs
quality, and quality is what the eval suite exists to measure.

Boundary

A test fails when any module outside firenze.model imports a provider SDK. It
caught two files on its first run — one false positive (naming a provider as a
config value is not coupling, so it now looks for the import) and one real (a
config default quietly naming a vendor's model).

85 passed · ruff clean · mypy strict: 20 files, no issues

Provider defaults to none until the pilot opens; fake covers development.