fix(dashboard): resolve model-switch spinner + persist benchmark on switchboard hosts

[azilber]

Problem

On switchboard-mode hosts (e.g. CPU installs after PR #288), the Models dashboard misbehaves:

1. "Working" spinner never resolves on model switch. The host agent records activation completion in data/model-state.json (record_verified_route) and never writes the legacy data/model-activation-receipt.json. The dashboard-api's activationReadyModel signal was gated solely on that receipt, so it stayed null forever and the UI (activationMatches in useModels.js) spun to its 610s timeout on every switch — even though the switch had actually succeeded (llama-server + model-state.json confirm it).
2. Benchmark result vanished after a few seconds. The UI discarded the benchmark response and re-derived Speed server-side on each poll, so the freshly measured tok/s only survived until the next refresh.

Fix

• dashboard-api (routers/models.py): add _read_model_state() + _switchboard_activation_complete() and OR it into the activationReadyModel gate. The legacy receipt path is preserved, so non-switchboard installs are unaffected. Reuses _model_name_tokens().
• dashboard UI (useModels.js, Models.jsx): capture the benchmark tokensPerSecond into a client-side benchmarkResults map so the value survives the next /api/models poll; getSpeedDisplay(model, benchmarkOverride) prefers it, threaded through the rows and the current-model panel. Poll acceleration also covers a running benchmark.

Not changed (working as designed)

• "Agent blocked" on Qwen 2.5 0.5B — catalog flags it not_agent_viable.
• Hermes "model doesn't update" — llama.cpp ignores the requested model name and serves whatever GGUF is loaded.

Tests

• Backend: 6 new switchboard-completion cases (positive + negative) in test_models.py; full test_models.py (114), test_model_state.py/test_config.py (99) pass.
• Frontend: new useModels test that benchmarked tok/s survives a later poll; ESLint clean; useModels (35) + Models (36) tests pass; vite build succeeds.

Verified live

Rebuilt/recreated ods-dashboard-api + ods-dashboard; with no receipt on disk, GET /api/models now returns activationReadyModel: "qwen3.5-9b-q4" (was null), sourced from the switchboard state.

Two pre-existing large files (models.py, tests/test_models.py, already >2000 lines) carry a # FILE-SIZE-OK marker; splitting is a separate follow-up.

🤖 Generated with Claude Code