Add CheaperInference provider - #7010

Action items

• [high] [violation] providers/cheaperinference/models/claude-sonnet-5.toml:98 - Check: Every toggle must have a leading top-of-file wire comment with the exact request path. Why: This multi-model OpenAI-compatible gateway adds toggle on five models (claude-sonnet-5, deepseek-v4-flash-0731, deepseek-v4.1-flash, kimi-k3, qwen-3-8-max) with only pricing comments above the first key. Callers and the catalog cannot tell which field turns reasoning on/off on this host (e.g. thinking.type, enable_thinking, reasoning.enabled). Action: Add a leading # Toggle: … (and matching effort wire path if applicable) on each of those five files documenting this host’s real control, or drop toggle if the gateway does not expose a separate on/off control.

Action items

• [high] [possible mistake] providers/cheaperinference/models/claude-sonnet-5.toml:5 - Check: Relay reasoning_options must match lab + same-surface peers for controls this host actually forwards. Why: Patch 2 removed toggle on claude-sonnet-5, both DeepSeek Flash models, kimi-k3, and qwen-3-8-max while stating the gateway forwards reasoning to the upstream provider. Lab/OpenRouter baselines for those models use a separate on/off control plus graded effort (no none in effort). Effort-only leaves no documented off path and understates capability if thinking / enable_thinking / equivalent fields pass through. Action: Either restore toggle (with a leading wire-path comment) on those five models to match lab/peers, or cite host docs/tests showing on/off fields are stripped or rejected and keep effort-only.
• [medium] [possible mistake] providers/cheaperinference/models/deepseek-v4-flash-0731.toml:17 - Check: Provider [limit] overrides must be real host deltas from the catalog API, not copy-paste caps. Why: Lab deepseek-v4-flash-0731 output is 384_000, but this file sets output = 65536 while sibling deepseek-v4.1-flash only bumps context and still inherits 384_000 output. That split is inconsistent for one gateway and matches common 64k defaults more than the lab Flash ceiling. Action: Confirm both DeepSeek limits from GET /v1/models (or equivalent) and align output/context with what the host actually serves; drop non-delta overrides.
• [low] [possible mistake] providers/cheaperinference/provider.toml:1 - Check: Rich provider catalog APIs should get a sync module when they can authoritatively drive model/pricing data. Why: The PR already keys rates off versioned GET /v1/models (pricing_version), so hand-maintained costs/limits will drift. Action: Add a sync module per sync.md, or note in the PR why this catalog stays hand-authored.

Addressed all three.

1. toggle: tested on the live /v1/chat/completions. reasoning: {enabled: false} removes reasoning tokens and the reasoning field on kimi-k3, qwen-3-8-max, deepseek-v4.1-flash and deepseek-v4-flash-0731, and enabled: true restores them (kimi-k3: 26 reasoning tokens → 0 → 26). Those four now carry toggle with # Toggle: reasoning.enabled = true|false as a leading comment. On claude-sonnet-5 neither reasoning.enabled nor thinking had any observable effect on this path, so it stays effort-only.
2. DeepSeek limits: both come from GET /v1/models. deepseek-v4-flash-0731 returns context_length: 1048576, max_output_tokens: 65536; deepseek-v4.1-flash returns context_length: 1048576, max_output_tokens: 384000. Hence 0731 overrides both and 4.1-flash only context.
3. Sync module: added to this PR (packages/core/src/sync/providers/cheaperinference.ts with tests, registration in sync/index.ts, and a CHEAPERINFERENCE_API_KEY line in the workflow). It only reads GET /v1/models, so a read-only key works; without the secret the matrix job for this provider fails the same way other keyed providers do. Replaying the live catalog through bun models:sync cheaperinference --dry-run gives 0 created, 0 updated, 0 removed, 16 unchanged against the files in this PR.

No actionable findings.