[frontend] Loading state with phase indicator (replace generic spinner)

[gapilongo]

Problem

Per-turn latency is 3-30 seconds (playbook-heavy turns + regen attempts push this higher). The current loading UX is presumably a generic spinner — the operator has no signal of what's happening, which feels broken when a 25-second turn drags on.

The backend pipeline has clear phase boundaries that we can surface:

1. Routing (intent + knowability + scope filter) — ~50ms
2. Substrate fetch (catalog match + playbook match + RAG retrieve + tool calls) — 100-2000ms
3. Model generation (vLLM streaming) — 2-20s
4. Verification (provenance verifier runs) — ~10ms
5. Regen attempt (if hard_failures) — repeats phases 3+4

A phase indicator turns dead time into useful signal.

Acceptance criteria

• Build a <PhaseIndicator> component shown in place of the current spinner:
  • 4-5 phase steps with icons (lucide-react: Route, Database, Cpu, ShieldCheck)
  • Current phase highlighted, completed phases checked, pending phases dimmed
  • On regen, the indicator restarts from "Model generation" with a small "(attempt 2/3)" badge
• Phase transitions driven by either:
  • Option A (simpler): time-based pseudo-progression — advance phases on a heuristic timer; not accurate but better than a spinner
  • Option B (proper, preferred): backend SSE / WebSocket emits phase events; frontend listens. Requires backend change — coordinate with backend folks before going down this path.
• If Option B requires backend work, scope this issue to Option A and open a separate issue for the SSE plumbing
• Should be visually integrated with the chat message bubble (not a modal)
• Reduce-motion friendly (prefers-reduced-motion: reduce → no animations, just static state)

Constraints

• No new dependencies — use existing lucide-react + Tailwind
• Keep it understated — this is operator UX, not a marketing animation
• Accessible — aria-live="polite" so screen readers announce phase transitions

Pointers

• Current chat component: frontend/src/components/chat/ (find the message-pending render)
• vLLM streaming is supported on the backend (see src/agent/model_adapter.py streaming path) — check whether the case_server currently relays it
• Phase definitions on the backend: walk src/agent/skills.py → _skill_answer_freeform for the canonical phase order

Difficulty

S (2-3 hours) for Option A (time-based heuristic), L (1-2 days) for Option B (real backend streaming). Recommend starting with A.