docs: add dedicated "Obtain models" guide for the pick-free install

[thinmintdev]

What

Adds a single dedicated documentation page, docs/guides/obtain-models.mdx (slug /docs/guides/obtain-models/, sidebar order 15), covering how an operator obtains models on hal0.

This addresses the model-obtainment documentation gap tracked in #1083 / #1084 (referenced, not closed). Since PR #1085 (60444fa/fc50839) the installer no longer ships model picks — a fresh box boots with services up but zero models and zero slots (hal0 setup --auto --no-slots seeds sentinel + wiring only). The page is written for that pick-free reality and never claims the installer selects or pulls a model.

Coverage

Documented with Dashboard and CLI side by side (Starlight Tabs):

1. Overview — the local registry as source of truth; the fresh pick-free starting state; NPU/FLM called out as the registry exception.
2. Model storage directory — default /var/lib/hal0/models; hal0 setup --storage-dir; [models].store (single source of truth) superseding the deprecated [models].pull_root; HAL0_MODEL_STORE; the /mnt/ai-models mount default.
3. Ways to add a model — pull from HF (curated alias / registered id / HF coords); register a local file; scan a directory; upstream-advertised models.
4. Initial pulls — the list → pull → assign → start-slot flow.
5. NPU / FastFlowLM — family:size tags, host flm cache, npu slot wiring.
6. Capabilities (ComfyUI) — capability-based image/video fetching.
7. Verifying / troubleshooting — pull status, where files land, gated-model HF token, disk preflight, missing-model slot fallback.

Cross-links to/from getting-started/first-model, guides/choose-models, guides/pull-and-register-models, guides/manage-slots, guides/generate-images, reference/cli, reference/paths-and-files. Leans on pull-and-register-models for the deep pull-lifecycle mechanics rather than duplicating them.

Real commands / endpoints documented (verified against source)

• CLI (src/hal0/cli/model_commands.py): hal0 model list [--json], pull <alias|id>, register <id> --path --name --license, rm, show, assign --slot. Slot start: hal0 slot load <name>.
• Pull: POST /api/models/<id>/pull (+ hf_repo/hf_filename/labels for by-coords); /pull/status. HF proxy: GET /api/hf/search, POST /api/models/inspect.
• Scan (verified there is no hal0 model scan CLI — API only, flagged explicitly): POST /api/models/scan/preview then POST /api/models/add-from-path per file; legacy full-root POST /api/models/scan (empty body), auto-run at startup when [models].auto_scan_on_start.
• Register local file: hal0 model register → POST /api/models; capability-detecting POST /api/models/add-from-path.
• NPU/FLM: hal0 model pull <family:size> auto-routes through host flm pull (is_flm_tag); FLM weights live in ~/.config/flm/models, not the registry.
• ComfyUI: POST /api/comfyui/models/fetch ({"auto":true} or {"selections":[...]}) with the CAPABILITIES registry (txt2img/img2img/txt2video/img2video/image_upscale).

Omitted / flagged

• No hal0 model scan subcommand exists — stated plainly; scanning documented via the real API + dashboard modal.
• No CLI flag for by-HF-coords pull — documented via the API body (as the dashboard's "Add by HF coords" modal uses).
• Corrected an initial draft error: hal0 slot create has no --device npu flag (--hardware is vulkan/rocm/cpu only) — NPU slot provisioning is routed through hal0 setup.

Verified

• astro build in the docs workspace (hal0-web) builds the page cleanly — /docs/guides/obtain-models/index.html generated, no MDX errors, no broken-link warnings.
• Frontmatter + Starlight component usage matches sibling guide pages; all internal links point at real slugs.

🤖 Generated with Claude Code