feat(aixy): expand gateway catalog (2/2)

[oscarcpozas]

Summary

• add the remaining 257 provider-qualified Aixy routes generated by the sync module in feat(aixy): add curated 50-model catalog #5506
• cover Alibaba, Azure, Bedrock, Cerebras, DeepSeek, Fireworks, Groq, Meta, Mistral, NVIDIA, Perplexity, Together, Vertex, and xAI
• factor every route against canonical Models.dev metadata while preserving provider-specific pricing, limits, modalities, lifecycle status, and reasoning controls

This is part 2 of 2. The split exists only to keep each PR below GitHub’s 300-file diff limit used by the repository reviewer. Together, #5506 and this PR expand Aixy from 1 to 506 models.

Catalog policy

Aixy is a BYOK gateway, so actual availability remains tenant-specific. The catalog lists reviewed first-class integration routes using the public provider/model contract. It excludes non-chat surfaces, tenant-defined community endpoints, and source routes without canonical model metadata.

Validation

• bun validate
• 257 model additions; 258 valid Aixy models in this half when combined with the existing route
• complete two-part catalog: 506 resolved Aixy models

Related

• Part 1 and sync module: feat(aixy): add curated 50-model catalog #5506
• Original Aixy provider: feat(aixy): add gateway provider #5462
• https://docs.aixy-gateway.com/integrations/overview

[github-actions]

Action items

• [high] [violation] providers/aixy/models/alibaba/qwen-flash.toml:2 - Check: Every toggle must have a leading top-of-file wire-path comment (AGENTS.md → Reasoning options; audit skill §3). Why: Dozens of new Aixy reasoners set type = "toggle" with only a generic “mirrors providers/…” note (or no toggle docs at all)—e.g. Alibaba Qwen routes, Bedrock/Azure Claude Sonnet 5, DeepSeek/Fireworks/Together/NVIDIA/Vertex Gemini 2.5 paths—so callers cannot see the exact request field Aixy/upstream expects (enable_thinking, thinking.type, chat_template_kwargs.enable_thinking, system-prompt toggles, etc.). Empty/[] is not the issue; undocumented toggles are. Action: Add a leading # Toggle: <exact wire path> (and effort/budget lines when present) on every new file that includes type = "toggle", not only a peer-file pointer.
• [medium] [possible mistake] providers/aixy/models/azure/kimi-k2.5.toml:1 - Check: Authored reasoning controls must match host evidence; comments must not contradict reasoning_options. Why: Header claims Azure’s chat-completions schema has “no toggle, effort, or reasoning budget,” yet the file sets [[reasoning_options]] type = "toggle" (same pattern in kimi-k2.6.toml). That is internally inconsistent and leaves consumers unsure whether Kimi reasoning is controllable on this Aixy/Azure route. Action: Either document the real toggle wire path and drop the “no toggle” claim, or set reasoning_options = [] if Azure truly exposes no control—and align kimi-k2.6 the same way.
• [medium] [possible mistake] providers/aixy/models/vertex/google/gemini-2.5-flash-lite.toml:3 - Check: Gateway routes that claim to mirror an upstream host must carry the same real capability deltas (incl. base_model_omit). Why: Upstream providers/google-vertex omits lab structured_output for gemini-2.5-flash-lite, gemini-2.5-pro, and gemini-flash-lite-latest, but the Aixy Vertex copies only set cost/reasoning_options and still inherit structured_output = true from models/google/…. That overstates structured output on those Aixy Vertex routes relative to the Vertex entries they mirror. Action: Add the same base_model_omit = ["structured_output"] (or an explicit override) on the matching Aixy Vertex files so resolved capabilities match Vertex.

[github-actions]

No actionable findings.

[rekram1-node]

this pr is too big

[oscarcpozas]

Consolidating this work into #5506 with a much smaller curated set based on the review feedback. Closing this bulk second half.