Report why a node's model inventory is empty

[DaveHomeAssist]

Problem

"Models aren't loading" was undiagnosable from the router. get_node_models collapsed every outcome into an empty list, so four distinct conditions reached the UI as the same "No models available":

• Ollama unreachable (connection refused)
• Ollama bound to loopback only, so reachable from the node but not from the router
• /api/tags slower than the hardcoded 3s timeout
• Node reachable with genuinely nothing pulled

GET /nodes/{node_id}/models returned HTTP 200 with models: [] in all four cases, and the only trace of the real cause was a print to the router's stdout.

Changes

Surface the failure reason. fetch_node_models returns {"models", "error"}. GET /nodes/{node_id}/models and GET /nodes/status both carry that error field. describe_node_error names the actual cause and, for a refused connection, points at the usual culprit (OLLAMA_HOST=0.0.0.0:11434). error: null with an empty list now unambiguously means "reachable, nothing pulled".

Malformed DAVE_NODES no longer fails silently. A parse failure registered zero nodes through a bare except Exception, which is indistinguishable from an empty cluster. It now prints the parse error and the expected JSON shape.

Timeout is configurable. DAVE_NODE_TIMEOUT (default 10s) replaces the hardcoded timeout=3 on both /api/tags calls. 3s is tight for a node loading a large model.

Stop clobbering a good inventory. The endpoint overwrote MODEL_INVENTORY[node_id] even when the fetch failed, so one transient timeout made /chat reject models the node really serves with Model '<id>' is not available. The inventory is now only written on a successful read.

Vision heuristic token boundaries. "mm" in lid matched every gemma tag and "vl" in lid was similarly loose, so ordinary text models were labelled vision-capable during load. Both now match as whole tokens, with an added pattern for the qwen2.5vl style.

Frontend. The model dropdown distinguishes "Node unreachable" (with the reason as a tooltip and in the console) from "No models pulled on node". loadModelsBtn.onclick = loadModelsFromNode passed the click MouseEvent in as preferredModelId, which failed every state.modelMeta lookup and silently reset the user's model choice on each manual reload; it now calls with no argument.

Verification

python -m py_compile app.py
python -m pytest -q                      # 31 passed
node --check static/app.js static/prompt-contract.js desktop/main.js desktop/preload.js
bash -n deploy/check-cluster.sh scripts/verify-cluster.sh
npm ci && npm ls --depth=0
git diff --check

Six tests added to tests/test_api_contracts.py covering the unreachable-node reason, inventory preservation across a failed fetch, the reachable-but-empty case, HTTP error surfacing, offline reason in /nodes/status, vision token boundaries, and the loud DAVE_NODES parse failure.

Cluster reachability itself is still not proven by repository tests — these changes make the runtime report which of the four conditions is actually happening, they do not verify a live node.

---

Generated by Claude Code