[Phase A1] Refactor interview UI to chat-style transcript with editable past answers

[terisuke]

Tracked under ADR 0002 — Phase A. Implementation detail: docs/implementation-plans/multi-persona-multi-format.md.

Problem

The current interview surfaces one question at a time but renders the conversation inside a max-height: 340px scroll-locked div (static/css/style.css:273). Past answers are not clickable, not editable, and not visually grouped with their parent fixed question. There is no way to fork the session if the user realises an earlier answer should be revised — the only escape is to start a fresh session, losing accumulated context.

This blocks the "feels like Claude/ChatGPT" comparison the project owner wants.

Scope

• Replace the bounded question-log div with a full-height scrolling transcript pinned to the centre column.
• Render every BriefAnswer as a clickable bubble; clicking opens an inline edit affordance.
• Editing an answer creates a new session that forks at that answer (server: POST /api/sessions/{id}/answers/{answer_id}/edit, returns the new session_id and parent_session_id). The original session is retained.
• Deep-dive question bubbles render visually nested under their parent fixed question, with the parent answer quoted as context.
• Keyboard: ↑ recalls the previous unsubmitted answer; Cmd/Ctrl+Enter submits.
• Question bubbles left-aligned, answers right-aligned (familiar chat metaphor).

Acceptance criteria

• Editing answer feat: CI/CDパイプラインの設定を追加 #2 in a 5-answer session produces a new session whose answer list is [1, 2', …].
• The original session is reachable from the new session via parent_session_id.
• Keyboard shortcuts work without breaking textarea newline insertion.
• Deep-dive bubbles visually attach to their parent fixed question.
• Existing scenario tests still pass; no domain or application layer changes required for this issue.

Out of scope

• Streaming (covered by Phase A2).
• Editable draft preview (covered by Phase A3).
• Persona switching (Phase B / C).

Files likely touched

• static/index.html
• static/js/script.js
• static/css/style.css
• internal/handlers/workflow.go (new fork-on-edit endpoint)
• internal/application/brief/service.go (fork-on-edit logic)

[terisuke]

2026-05-02 planning update: keep this in Phase A, but implement it after #18.

The transcript should reuse the streaming primitives from #18 instead of introducing another spinner/blocking path. Scope remains chat-style transcript, editable answer bubbles, and in-memory fork-on-edit. Durable fork history still belongs to Phase C / #26.

[terisuke]

Prerequisite runtime work is now merged in PR #46 before starting #17. The LLM path is now Evo X2 Ollama primary -> Evo X2 llama.cpp fallback -> workstation-local llama.cpp last resort, with phase defaults set to gemma4:e2b for source/style summarization, qwen3.6:27b for follow-up questions, and gemma4:31b for final Japanese drafts. Future llama.cpp swap orchestration is tracked separately in #45, so #17 can proceed on top of this stable routing.

[terisuke]

Implemented in PR #49 and merged into develop. The interview UI now renders a chat-style transcript, answer bubbles support inline edit/fork, the fork response returns parent_session_id, keyboard shortcuts are wired, and docs/ADR status were updated. Validated with go test ./..., node --check static/js/script.js, and local server smoke checks.