Feature Request: First-class support for preset routers, split endpoints, and serverless inference providers

Is your feature request related to a problem? Please describe.

Hermes's current custom_providers system works well for standard OpenAI-compatible endpoints but has friction with several modern inference patterns:

1. OpenRouter Presets — Private @preset/* models fail model-discovery warnings and pollute the model list with 300+ public models unless discover_models: false is manually added (undocumented for this use case).

2. Split Endpoints (same base_url, different keys) — Providers like Wafer AI have multiple product tiers (pass, serverless) sharing the same base_url but requiring different API keys. The current deduplication logic (seen_name_url_pairs) silently drops duplicate base_urls, forcing users to inline keys in model_aliases.

3. NVIDIA NIM — No native provider profile; users must use custom with base_url: https://integrate.api.nvidia.com/v1. Model catalog discovery works but lacks NIM-specific features (e.g., nvidia/ namespace handling).

4. Nebius Token Factory — Token-based auth (long-lived service tokens) works but isn't clearly distinguished from standard API keys in docs.

5. Modal Direct — No built-in support for Modal's direct inference endpoints (https://api.<region>.modal.direct/v1/chat/completions).

Describe the solution you'd like

1. Native openrouter_presets provider type

custom_providers:
  - name: my-openrouter-presets
    type: openrouter_presets   # new type
    api_key: sk-or-v1-...
    presets:
      - "@preset/some-name"
      - "@preset/some-other-name"

• Automatically sets discover_models: false
• Validates preset slug format (@preset/*)
• Suppresses "not found in model listing" warnings for preset models

2. Explicit provider_id for split endpoints

custom_providers:
  - name: wafer-pass
    provider_id: wafer-pass    # new field
    base_url: https://pass.wafer.ai/v1
    api_key: wfr_...
  - name: wafer-serverless
    provider_id: wafer-serverless
    base_url: https://pass.wafer.ai/v1
    api_key: wfr_...

• Deduplication uses (provider_id, base_url) instead of just (name, base_url, model)
• Allows same base_url with different keys

3. Native nvidia_nim provider

providers:
  nvidia_nim:
    api_key: nvapi-...
    default_model: deepseek

• Built-in profile with correct models_url, headers, and nvidia/ namespace handling

4. Native modal provider

providers:
  modal:
    api_key: modalresearch_...
    base_url: https://api.us-west-2.modal.direct/v1
    default_model: zai-org/GLM-5.1-FP8

5. Better discover_models documentation

Document discover_models: false for:

• Private endpoints
• Preset-only providers
• Dedicated single-model endpoints

Describe alternatives you've considered

Current workarounds:

• OpenRouter presets: discover_models: false + custom provider (works but undocumented)
• Split endpoints: Inline api_key in model_aliases (fragile, repeats key)
• NIM/Modal: Generic custom provider (loses type-specific optimizations)

Additional context

Our production custom_providers config for reference:

custom_providers:
  - name: wafer-pass
    base_url: https://pass.wafer.ai/v1
    api_key: wfr_...
    model: GLM-5.1
  - name: wafer-serverless
    base_url: https://pass.wafer.ai/v1
    api_key: wfr_...
    model: deepseek-v4-flash
  - name: Openrouter Presets
    base_url: https://openrouter.ai/api/v1
    api_key: sk-or-v1-...
    discover_models: false
    model: '@preset/some-name'

Checklist for maintainers

• openrouter_presets provider type
• provider_id field for split endpoints
• nvidia_nim built-in provider
• modal built-in provider
• Document discover_models: false patterns