fix(chat): persist agent plan in chat thread across reload

[AVADSA25]

User-visible bug

Drop a project in Project mode → CODEC drafts a plan and renders an inline card with Approve / Reject / View-plan buttons → refresh the chat OR click another session → card is gone, only the original user message + a "blocked_on_permission" status pill remain.

The user reported this multiple times. Blocked the forex anchor demo on 2026-05-03.

Root cause (3 stacked bugs)

1. codec_chat.html:819 — pushed only "Project drafted: agent_xxx" to local chatHist and didn't call saveMessages() at all for the project flow. The actual plan card with all the info lived in the DOM only.
2. Even if saved, addMessage() renders plain text — no re-render path knew the message was a project drop.
3. startNewSession() ran on every page load, generating a fresh sessionId. Hard refresh orphaned the previous chat under an id only reachable via the sidebar.

Fix

1. Embed a marker [CODEC_AGENT_PLAN:<id>] in the saved assistant message content. Pattern locked to the JS regex:

   var AGENT_PLAN_MARKER_RE = /\[CODEC_AGENT_PLAN:(agent_[a-z0-9]+)\]/;

2. Save the assistant message via /api/qchat/save (was missing).
3. Refactor inline HTML into renderAgentPlanCard(id, info) that handles BOTH response shapes:
   • POST /api/agents flat: {agent_id, status, project_dir}
   • GET /api/agents/<id> nested: {manifest, plan, state, grants}
     Status-aware buttons:
   • pending → Approve + Reject + View
   • blocked_* → View+resolve + Abort
   • terminal (done/aborted/plan_failed) → View only
4. New rehydrateAgentPlanCards() scans #messages .msg.assistant .msg-bubble after loadSession(), extracts agent_ids, fetches in parallel, replaces bubbles with live cards. On failure, shows "agent state not found — manifest may have been removed."
5. Persist sessionId in localStorage('codec-chat-session'). On boot, if a saved session exists and is non-empty, replay it through the same code path the sidebar uses.

Tests

tests/test_chat_plan_persistence.py — 6/6 pass:

• extractAgentIdFromMessage (Python parity port of the JS regex) finds real markers
• Returns empty for no-marker / null / malformed
• Ignores malformed markers (wrong case, missing brackets, no agent_ prefix)
• Byte-for-byte SQLite save→load round-trip preserves the marker
• Multi-marker session preserves ordering
• Locked to real-world hex id format (agent_1416ea3e1b02), rejects underscored "ids"

Hermetic: re-creates qchat schema in a temp sqlite, exercises the same INSERT/SELECT codec_dashboard.qchat_save/qchat_session use. Avoids the codec_dashboard import (pynput chain).

Test plan (manual after merge)

• Drop a project in Project mode → card renders ✓ (existing behavior)
• Click another session in the sidebar → return → card still there
• Hard refresh (cmd+R) → land back on same session → card still there
• Approve the plan → card updates to show running status (status-aware re-render)
• Block the agent on a permission → card shows "View+resolve" + "Abort" buttons
• Delete the manifest manually → card shows "agent state not found" notice instead of disappearing

🤖 Generated with Claude Code