fix(provider): add Cheaper Inference as a built-in provider

[aiapienthusiast]

Adds Cheaper Inference as a built-in provider.

Cheaper Inference is an OpenAI-compatible gateway. One endpoint and one API key reach models from several upstream providers. The entry declares "response_type": "OpenAI" and uses the existing OpenAI request path, so this change needs no new response handling.

The change follows #3623 (fix(provider): add OrcaRouter as a built-in provider), which added OrcaRouter in the same way.

Provider details

The repository has a provider-integration issue template that asks for these details:

Field	Value
Provider name	Cheaper Inference
API documentation	https://cheaperinference.com/docs
API compatibility	OpenAI-compatible
Authentication method	API key, read from CHEAPERINFERENCE_API_KEY
Chat completions	https://api.cheaperinference.com/v1/chat/completions
Model list	https://api.cheaperinference.com/v1/models

Changes

File	Change
crates/forge_repo/src/provider/provider.json	New cheaper_inference entry: API key variable, OpenAI response type, chat-completions URL, models URL, api_key auth.
crates/forge_domain/src/provider.rs	ProviderId::CHEAPER_INFERENCE, added to built_in_providers() and to the FromStr match. Plus a display_name arm, because the default conversion gives CheaperInference and the display name is Cheaper Inference. Three tests mirror test_orca_router_*. The new id joins the two shared assertion lists.
crates/forge_repo/src/provider/provider_repo.rs	test_cheaper_inference_config, which mirrors test_orca_router_config.
README.md	Provider env-var block, in the same shape as the surrounding <details> sections.

The model list stays dynamic

provider.json does not enumerate the models. The entry points models at the provider's /models endpoint. The catalogue and the per-token pricing are discovered at runtime, not pinned in the repository. dto::openai::Pricing already reads that pricing and treats it as optional.

#3623 did the same. test_cheaper_inference_config asserts the Models::Url variant, so a later switch to a hardcoded list fails loudly.

What this PR does not do

This PR does not touch crates/forge_app/src/dto/openai/model.rs and adds no /models fixture.

In #3623 that part was a schema relaxation, which OrcaRouter's response shape made necessary. Those fields are already optional on main. This PR changes no deserialisation. No response from this provider is captured, so a fixture would be invented, not recorded.

Testing

Run against 7ef8386acaa9564a76526828c53d430e1696730a, the current main:

• cargo test --lib -p forge_domain — 618 passed, 0 failed, with the three new test_cheaper_inference_* cases
• cargo test --lib -p forge_repo — 334 passed, 0 failed, 1 ignored, with the new test_cheaper_inference_config
• cargo clippy -p forge_domain -p forge_repo --lib --all-targets — clean, 0 warnings
• rustfmt --check on the two touched Rust files — no diff

Toolchain limit

The preparation environment cannot install the toolchain that rust-toolchain.toml pins (1.97), because static.rust-lang.org is unreachable there. The checks above ran on stable 1.95. Two points need a maintainer's attention:

• rustfmt ran without the nightly-only options in .rustfmt.toml. Stable rustfmt ignores them with a warning. The new code adds no imports and wraps to 80 columns, which is what those options govern. Please re-run cargo fmt on the pinned toolchain.
• cargo clippy ran clean on the two crates, but on the 1.95 lint set, and not across the whole workspace.

Not tested through forge

A direct check with a real key confirmed the endpoint. GET /v1/models returns the catalogue. POST /v1/chat/completions answers. The response carries the bare model id that was sent.

A reviewer should confirm the path through forge: the provider resolves from CHEAPERINFERENCE_API_KEY, and a completion succeeds from the app.