test(fireworks-ai): lock reasoning controls

[rekram1-node]

Summary

• audit every Fireworks provider catalog entry against the current Fireworks reasoning API reference, endpoint pages, serverless documentation, and Fire Pass documentation
• preserve exact positively supported controls for 7 entries and remove unverified toggle claims from 6 reasoning-enabled endpoints
• add an exhaustive catalog regression test that rejects unsupported options, options on non-reasoning entries, and unreviewed catalog additions or removals

Verified controls

• DeepSeek V4 Pro and Flash: toggle plus distinct effective high / max efforts
• GPT-OSS 20B and 120B: effort low / medium / high
• MiniMax M2.5 and M2.7: required reasoning with effort low / medium / high
• Qwen 3.6 Plus: toggle, effort low / medium / high, and integer reasoning budget with min = 1 and no documented maximum

For Qwen, Fireworks integer reasoning_effort accepts a positive integer and applies it as a hard cap on generated thinking tokens. This maps to { type = "budget_tokens", min = 1 }. The separate Anthropic-compatible thinking.budget_tokens request shape has a 1024 minimum, but that does not narrow Fireworks integer reasoning_effort.

Unresolved provider gaps

The following endpoints remain reasoning = true but intentionally omit reasoning_options. Fireworks documents reasoning behavior or availability, but its public endpoint-specific contract does not establish accepted disable semantics for these exact endpoints:

• accounts/fireworks/models/glm-5p1
• accounts/fireworks/models/kimi-k2p5
• accounts/fireworks/models/kimi-k2p6
• accounts/fireworks/routers/glm-5p1-fast
• accounts/fireworks/routers/kimi-k2p6-fast
• accounts/fireworks/routers/kimi-k2p6-turbo

No reasoning_options = [] or reasoning = false fallback was encoded because neither has positive evidence. No live tests were run because FIREWORKS_API_KEY is unavailable.

Validation

• bun validate
• bun test packages/core/test/fireworks-ai-catalog.test.ts (1 pass, 20 assertions)
• git diff --check