feat(api): let a model write the case, without letting it decide anything

[Madeuss]

Closes E1 (#5). The last piece of phase 1, and the first code in the repository
that calls a model.

The two boundaries

write() takes Case, never CaseWithSolution. Same discipline as the
solver, different reason: a writer who knows the culprit telegraphs them. Not
deliberately — the guilty character just ends up with the sharper description.

The test asserts the stronger property rather than the obvious one:

for suspect in case.suspects:
    assert user.count(suspect.name) == 1   # only in the cast list
    assert user.count(suspect.id) == 1

Nothing in the prompt distinguishes one suspect from another. A model that
cannot tell them apart cannot write the guilty one as guiltier.

The prompt only ever contains public facts. The veneer has no use for a
restricted fact, so it never receives one — and therefore cannot leak one. Which
means the canary check on the output is not pointed at the model. It is pointed
at us: a canary here proves context assembly upstream is broken (RN-010,
RN-012).

What the model is allowed to produce

A narrow schema: role_title, appearance, manner, and a scene. There is
no field for evidence, for a location, for a time, or for a culprit. Anything
outside those fields cannot reach the game because there is nowhere to put it.

What the schema can't prevent, validation rejects:

Check	Why it exists
canary in output	RN-012 — and it means an upstream bug, not a bad sentence
invented / dropped / duplicated character	the cast is not the model's to edit
one suspect described with another's name	survives a read-through, then contradicts every alibi
blank or runaway field	a long field is a model that started narrating the case

A rejected draft is discarded, never repaired. A model that broke the cast
list once will break it differently on a patch, and a half-corrected veneer is
harder to reason about than none.

Failure is not fatal

No key, a refusal, a timeout, a rejected draft — the case still prints, with a
note on stderr:

$ make case SEED=42 VENEER=1
veneer skipped: the model call failed: Could not resolve authentication method...
Caso 42 (gerador v2, pt-BR)
...

Prose is a luxury. The mystery is not.

Testing without spending anything

63 tests, none of which need an API key. The model is a stub that returns
exactly the drafts that must not reach a player: the invented character, the
swapped name, the leaked canary, the refusal. Waiting for a real model to emit
one of those by chance is not a test strategy.

Two decisions worth flagging

Model. claude-opus-5, switchable through MANSAO_VENEER_MODEL because it
is the one knob that moves cost per case. I have not measured the cost — there
is no key on this machine — but it is one call per case, and adaptive thinking
means output tokens dominate. Worth measuring against the R$ 0,50 per match
budget before phase 2 adds six NPCs with a turn each.

No server-side refusal fallback. The usual default for Opus 5 code is to
pass fallbacks, so a policy decline reruns on another model inside the same
call. Skipped here deliberately: the veneer is optional, so degrading to no
prose is both simpler and more honest than silently switching models on a
creative-writing task. A refusal on a cosy murder mystery would be interesting
information, and swallowing it would hide that.

Also in this PR

The ADR-0006 rename that phase 0 left behind — Configuracao → Settings, and
/health now answers version / environment. openapi.json regenerated.

63 passed · ruff clean · mypy strict: 16 files, no issues