feat(webui): expose a truthful model catalog and capabilities

[inureyes]

Parent / dependencies

Parent: #1834. Depends on #1835, #1839. Lifecycle fields integrate with the coordinator contract when available; no duplicate pool state.

Ownership

Own src/server/webui/catalog.rs, metadata/cache projections, capability mapping and catalog fixtures/tests. Reuse RouterSources, discover_models, downloader store/completeness and src/models/{registry,detection}.rs. Do not change GPU loaders or add a handwritten family support list.

Required behavior

• Enumerate configured model directory, managed cache and preset sources using existing precedence and containment. Handle same basename/different roots, aliases, preset overrides, hidden/deduplicated entries, missing default store, unreadable/corrupt entries and partially installed snapshots. Keep one catalog identity and a separate actual inference ID.
• Return deterministic pagination/search/filter (50 default, 200 maximum), source and removal eligibility, exact model_type/architecture, task/modalities, quantization and disk bytes. Distinguish complete, supported architecture, supported backend/task, tested checkpoint evidence if available, and runtime ready. Unknown metadata is null with a reason. Family heading match is not loadability.
• Derive static capabilities from registry/detection plus checkpoint metadata and runtime gates; after load merge provider-confirmed capabilities without pretending static support is a measured successful load. Unsupported/experimental entries remain inspectable but unusable actions explain why. Non-chat families get task/API guidance, not a chat launch.
• Read bounded metadata on blocking workers; do not read safetensor payloads, initialize tokenizers/models, follow out-of-root symlinks or scan the entire home directory. Disk-size traversal is bounded/background/cached, not repeated every status poll. Canonicalize/revalidate at the operation boundary as well as discovery.
• POST /catalog/refresh is an explicit bounded rescan operation. Merely viewing the library or bootstrap causes no writes/downloads/preloads. Show configured discovery roots with safe display paths; there is no HTTP filesystem picker. Support the repository store via explicit --models-dir models/mlx, not a cwd-dependent default.

Acceptance criteria

• Fixtures cover managed owner/name cache plus flat models/mlx root, duplicate names, symlink escape/loop, unreadable directory, malformed/large config, incomplete shard set, unsupported dense variant under a supported vendor, non-chat task and backend unavailable.
• Empty library works; metadata-only tests assert zero network/provider/tokenizer creation and bounded IO. A 1,000-entry synthetic inventory paginates stably through refresh without duplicate/missing IDs.
• Capability tests compare against the registry and representative provider readiness; no fabricated memory estimate or zero used for unavailable information.
• Catalog integration verifies single-model mode reports its actual model and correct read-only reasons without registering a second provider.