feat(pleumrouter): add PleumRouter provider

[gachon-star-want]

What

Adds PleumRouter as a provider. PleumRouter is a Korea-region multi-provider LLM gateway with an OpenAI-compatible API (/v1/chat/completions, plus Anthropic Messages and OpenAI Responses adapters).

• Base URL: https://router.pleum.ai/v1
• Auth: PLEUMROUTER_API_KEY (plm_… keys)
• SDK: @ai-sdk/openai-compatible
• Docs: https://router.pleum.ai/docs

Who we are / why this PR

We are Pleum AI (pleum.ai). OpenCode and other agents fetch this registry at runtime — listing PleumRouter here is what makes us selectable in one click instead of manually pasting a custom base URL. Same pattern as other OpenAI-compatible gateways already in the registry (Requesty, etc.).

Follow-up to #2714 / #2722

Addresses maintainer feedback:

• Official Pleum mark as logo.svg (currentColor, no fixed dimensions)
• base_model inheritance for featured text models
• Explicit reasoning_options = [] on inherited reasoning models (no documented provider-specific reasoning controls)
• Costs = effective USD/1M (gateway markup included)

Happy to sync against the live /v1/models catalog or adjust model selection further.

Checklist

• Provider toml + logo
• Model tomls use base_model where possible
• reasoning_options explicit on reasoning models

Made with Cursor

[github-actions]

Action items

• [high] [violation] providers/pleumrouter/models/qwen3-max.toml:2 - Check: reasoning_options must not be set when the model is non-reasoning. Why: This file sets reasoning_options = [] but inherits reasoning = false from models/alibaba/qwen3-max.toml (line 7). The schema's refineModel rule rejects this with "Cannot set reasoning_options when reasoning is false", so bun validate will fail on this provider. All other PleumRouter models correctly inherit reasoning = true. Action: Remove the reasoning_options = [] line from qwen3-max.toml; non-reasoning inherited models must leave reasoning_options unset.
• [medium] [possible mistake] providers/pleumrouter/models/glm-5.1.toml:3-4 - Check: Provider cost overrides must match the actual model being served. Why: Nine of the ten new models follow a uniform 1.03× markup over the correct upstream price (e.g. claude-opus-4-8 5→5.15/25→25.75, gpt-5.5 5→5.15/30→30.9, qwen3-max 1.2→1.236/6→6.18, deepseek-v4-pro 0.435→0.44805/0.87→0.8961). glm-5.1 breaks this pattern: its 1.03/3.296 equals 1.03× the glm-5 price (1.00/3.20), not 1.03× the glm-5.1 price (Zhipu 1.4/4.4 → expected 1.442/4.532). This looks like a copy-paste from the wrong upstream model. Action: Verify and correct the glm-5.1 input/output cost against PleumRouter's actual price list.
• [medium] [possible mistake] providers/pleumrouter/models/*.toml (filenames) - Check: Registry model IDs (derived from filenames) must match the IDs the provider's API actually accepts. Why: Every comparable multi-provider OpenAI-compatible gateway in the registry (OpenRouter, Requesty, Kilo, FastRouter) exposes provider-prefixed IDs via nested folders (anthropic/claude-fable-5, openai/gpt-5.5, z-ai/glm-5.1, alibaba/qwen3-max), and opencode sends the registry model ID directly to the provider. These flat IDs (claude-fable-5, gpt-5.5, glm-5.1, etc.) will fail to resolve if PleumRouter's /v1/models returns prefixed IDs. Action: Confirm the exact IDs returned by PleumRouter's live /v1/models catalog and rename files into nested folders so filenames match the API model IDs.
• [low] [possible mistake] PR body - Check: Data-changing PRs should cite direct provider pricing/API sources mapped to the changed values. Why: The PR introduces per-model USD/1M costs and reasoning_options = [] claims but cites only a generic docs URL; nothing maps to the specific prices, and no API-reference citation supports the claim that no reasoning controls are forwarded (comparable gateways expose reasoning_effort/thinking/budget fields). These material factual values cannot be reviewed from the PR alone. Action: Add direct citations to PleumRouter's pricing page (what each cost value represents) and API reference (whether reasoning_effort, thinking, or budget fields are accepted/passed through).