feat(dashboard): AI Discuss panel + per-row Analyze button (#419)

[jonathaneoliver]

Summary

• The user-facing surface for AI session analysis. Floating 💬 Discuss panel on session-viewer.html (multi-turn chat about the loaded session) + per-row Analyze button on sessions.html (one-shot overview modal).
• Sub-issue 7 of epic epic: AI-assisted session analysis (chat + Analyze + compare) #412 — completes the MVP. Closes #419.
• Stacked on feat(forwarder): system prompt for session chat (#418) #430.

What's new

File	Role
content/shared/llm-chat.js	DOM-free SSE consumer + window.LLMChat API
content/dashboard/llm-discuss-panel.js	Discuss panel for session-viewer.html
content/dashboard/llm-analyze-modal.js	Analyze button + modal for sessions.html
content/dashboard/session-viewer.html	+2 script tags
content/dashboard/sessions.html	+2 script tags

Behaviour

Discuss panel (session-viewer.html):

• 💬 floating toggle bottom-right; slide-out right-side panel.
• Multi-turn chat tied to the URL's session=<id> parameter.
• localStorage persistence (last 50 messages per session).
• Model picker dropdown sourced from /api/llm_profiles.
• Budget meter above history; polls /api/llm_budget every 30s while open.
• Cmd/Ctrl+Enter sends; Escape closes (when not streaming); focus returns to toggle.
• Cited mm:ss.ms timestamps render as click-to-seek anchors that dispatch document.dispatchEvent(new CustomEvent('llm:seek')). Wiring to the chart cursor is a follow-up.

Analyze button (sessions.html):

• MutationObserver injects an Analyze button next to each row's star (the canonical [data-star-cell][data-session-id] hook).
• Click opens a modal that streams a one-shot overview from the same backend.
• z-index:3 sits above the row's stretched-link overlay so clicks don't navigate away.

Pre-commit code-review fixes (this is the meaningful section)

The Sonnet pass caught one critical bug: my initial implementation queried .session-card divs for the picker — but sessions.html actually renders rows as <tr> with the session_id on a child <span data-star-cell>. The button would never have appeared. Fixed to target the real DOM shape, with z-index above the row overlay.

Plus important polish:

• SSE parser now tolerates spec-conformant event:foo (no space) per W3C EventSource. Original required the literal event: (with space). Forwarder happens to emit the spaced form, but a future emit site or rewriting proxy would have silently dropped events.
• Profile-select race — Send button starts disabled ("Loading models…"); guards sendMessage until the profile fetch resolves. Stops a fast Cmd+Enter from sending with an empty profile.
• Modal Esc handler now removes itself on every close path (backdrop click, X button, Escape). Old code only removed on Escape — opening many modals would leak listeners.
• Discuss panel — Esc closes (when no stream is active); focus returns to the toggle button after close (WCAG 2.4.3).
• CSS namespace tightened: .llm-msg / .llm-ts → .llmchat-msg / .llmchat-ts to avoid future collision with generic class names.
• Duplicate id="llm-profile-select" removed — label now wraps the select so no id is needed (script-double-load can't create dupes).
• readOnly during streaming instead of disabled so users can still select / copy their own pending text.

Cross-compare per memory rule

testing-session.html (the LIVE control plane) is intentionally NOT touched. Analytical AI is for archived replay; the live-session control AI surface is sub-issues #424 (read-only inspection tools — get_session_state, get_failure_settings) and #425 (HITL mutation tools with propose/apply UX), both planned for after the analytical MVP ships.

What needs your hands-on browser verification before merge

(I'm a CLI agent — these are the gaps the reviewer + I flagged that only show up in a real browser.)

• 💬 toggle position vs existing bottom-right chart-expand controls on session-viewer.html — visual collision check.
• Per-row Analyze button placement in the picker's first cell — does it look right next to the star, does the click reach the modal (not the row navigation)?
• Real session_id end-to-end: pick an archived session, click Analyze, confirm a streamed answer arrives with click-to-seek timestamps.
• Open Discuss panel, ask a follow-up question, confirm history persists across page reload.
• Set LLM_DAILY_BUDGET_USD=0 and confirm the budget meter renders as exhausted + the chat refuses with a 429 message inline.
• Cancel mid-stream — does the Cancel button actually stop the upstream LLM call (server-side llm_calls.status='cancelled')?

Three files, all node --check syntax-clean. The forwarder side is fully covered by the existing 67-test suite (unchanged in this PR).

🤖 Generated with Claude Code

[jonathaneoliver]

Closing — superseded by the v2.0.0 release.

• The AI-chat backend (analytics: ClickHouse llm_reader user (read-only, hard server-side limits) #413–prompts: initial session_chat.md (schema + tool docs + citation rules) #418) was reimplemented and expanded directly on main: analytics/go-forwarder/llm_client.go, llm_session_chat.go, llm_profiles.go, llm_tool_query.go, llm_ledger.go, prompts/chat.md, plus the tiered tools + Anthropic client + chat persistence that didn't exist in this stack.
• The dashboard pieces (dashboard: Discuss panel + Analyze button + model picker on session-viewer.html / sessions.html #419–dashboard: two-session AI compare from sessions.html multi-select #421) targeted the legacy content/dashboard/ + content/shared/ surface, which the Vue 3 dashboard rewrite removed. Their diffs no longer apply to any file on main.

This stale stacked-PR chain (epic #412) has no salvageable diff against current main. The feature lives on in the v3 dashboard + forwarder; further work should target those, not this stack.