Custom build: model-role dashboard with dropdowns and UI polish

[f74dragon]

What changed

This contribution packages the local custom build into one coherent change:

• Plugin-local model-role dashboard (scripts/role-dashboard.py): a loopback-only web UI on 127.0.0.1:8765 that assigns a model identifier and reasoning effort to the four workflow roles (primary_orchestrator, native_implementer, native_reviewer, luna_task). Saving writes only config/role-map.json and regenerates the two plugin-local native-agent templates. It never queries, configures, or edits OpenCodex or any other application.
• Dropdown model selection: each role's model field is a dropdown populated from config/models.json (seeded with the Sol Advisor combos plus identifiers found in the local OpenCodex catalog and Codex Router), with a Custom model… option. Saving a custom model adds it to the dropdown for next time.
• Redesigned dashboard UI: the previous layout was oversized and clipped long identifiers. The new design uses a compact header, a calmer palette, a two-column card grid (single column on mobile), and monospace ellipsis-safe selects that render even the longest identifiers without clipping.
• Hardened installer flow: install-agents.sh --sync / --check, the runtime inspector, and an extended verify.sh (including a new model-list check) are part of the checkout.

Why

Typing model identifiers by hand was error-prone. The dropdown guarantees known, valid identifiers while still allowing custom entries, and saving a custom model grows the list automatically.

Impact

• Users get a local dashboard to configure roles without touching Codex or OpenCodex configuration.
• The native lane still requires the explicit companion install (install-agents.sh + --check) and a fresh Codex task; the Luna lane remains an explicit opt-in only.
• The dashboard validates only safe identifier syntax and known reasoning-effort values; it does not claim a provider accepts a selected value.

Validation

• sh plugins/sol-advisor/scripts/verify.sh passes end to end (role map, native templates, installer migrations, model list, native/Luna contracts, shell syntax).
• Dashboard exercised through its API (GET/POST /api/config) and headless-Chrome layout QA: no horizontal overflow at desktop or mobile widths, no clipping of the longest identifier, all dropdowns populate, and the Custom model… field reveals and focuses correctly.