feat: add model files endpoint

[fl0rianr]

Summary

Adds a backend-owned per-model file inventory endpoint for the GUI Files tab:

• GET /api/v1/models/{id}/files
• GET /v1/models/{id}/files

This is intentionally separate from storage/system metrics. Storage endpoints
answer drive/system usage questions; this endpoint answers which resolved local
files belong to one selected model.

Response shape

{
  "model_id": "Qwen3-4B",
  "files": [
    {
      "name": "model.gguf",
      "path": "/abs/path/model.gguf",
      "role": "main",
      "size_bytes": 123456789,
      "exists": true
    }
  ]
}

Implementation notes

• Adds ModelManager::list_model_files() using ModelInfo::resolved_paths as the first-slice source of truth.
• Uses std::filesystem::exists/existing safe filesystem helpers and the existing resolved-path size helper.
• Sums directories recursively, matching existing model-size behavior.
• Registers /models/{id}/files before the generic /models/(.+) route so foo/files is not interpreted as a model ID.
• Returns the public/wire model ID via the same alias resolution path as GET /v1/models/{id}.

Validation

PR 2 backend smoke test: /models/{id}/files works for Gemma-4-E2B-it-GGUF

Validated both public route variants:

GET http://localhost:13305/v1/models/Gemma-4-E2B-it-GGUF/files
GET http://localhost:13305/api/v1/models/Gemma-4-E2B-it-GGUF/files

Both endpoints return the same expected response shape:

{
  "model_id": "Gemma-4-E2B-it-GGUF",
  "files": [
    {
      "name": "gemma-4-E2B-it-Q4_K_M.gguf",
      "path": "/home/vscode/.cache/huggingface/hub/models--unsloth--gemma-4-E2B-it-GGUF/snapshots/ecc8b33b2c50598815e4b0f7cea6088e3ae7adb8/gemma-4-E2B-it-Q4_K_M.gguf",
      "role": "main",
      "size_bytes": 3106736256,
      "exists": true
    },
    {
      "name": "mmproj-F16.gguf",
      "path": "/home/vscode/.cache/huggingface/hub/models--unsloth--gemma-4-E2B-it-GGUF/snapshots/ecc8b33b2c50598815e4b0f7cea6088e3ae7adb8/mmproj-F16.gguf",
      "role": "mmproj",
      "size_bytes": 985654080,
      "exists": true
    }
  ]
}

Result:

main file is detected with role main
mmproj file is detected with role mmproj
both paths resolve to existing files
total listed size: 4092390336 bytes, roughly 3.81 GiB
/v1/... and /api/v1/... behave consistently
</details> ```

[bong-water-water-bong]

Two issues that should be addressed:

1. HIGH — Missing /api/v0 and /v0 route variants: Every other model endpoint follows the quad-prefix convention (/api/v0, /api/v1, /v0, /v1). This endpoint only registers under /api/v1 and /v1. Requests to /api/v0/models/foo/files will fall through to the generic model-by-id route and return "Model not found."

2. HIGH — Absolute filesystem paths exposed in API response: The existing model_info_to_json deliberately omits resolved_paths. This new endpoint exposes full paths like /home/user/.cache/huggingface/hub/models--..., revealing home directories, usernames, and cache layout. If absolute paths are truly needed by the GUI, at minimum document this in the API docs as intentional info disclosure.

Minor: TOCTOU gap between model_exists() check and list_model_files() call (pre-existing pattern, but worth wrapping in try-catch for 404). Missing #include <vector> in model_manager.h.

[jeremyfowers]

LGTM!