Unify live scanner ingest and fix Claude Desktop events

[Amorpheuz]

Summary

• route Claude, Cursor, and Codex live transcript scanning through the same canonical Ingest Event parsers used by transcript import
• centralize event preparation, wire adaptation, origin-scoped deduplication, usage folding, interruption handling, and live-shell fallback responses
• add Cursor plan support, whole-transcript subagent preparation, consecutive-response collapse, and interruption-during-tool coverage
• fix Claude Desktop lifecycle semantics: Stop ends a turn, SessionEnd closes the Session, and a later SessionStart reopens it
• keep Claude Desktop's internal follow-up suggestion subagents out of the timeline while retaining them as ignored raw-ingest evidence, including migration repair for existing Sessions
• keep hook-data-wins reconciliation covered at the Collector seam

Why

Live hooks and transcript import previously encoded transcript behavior independently. That duplication caused response loss, repeated Cursor commentary, and Origin drift. Claude Desktop also exposed two existing Collector assumptions: turn-level Stop hooks were treated as Session termination, and internal suggestion generators appeared as user subagents.

This change makes each per-agent parser and the canonical adapter authoritative for both Origins, while keeping trigger-only behavior in thin live shells. Collector lifecycle classification is now shared by display and persistence.

Developer impact

• transcript-format and payload fixes land once for live and imported Sessions
• live scanner Events use origin-namespaced explicit keys without changing hook-data-wins reconciliation
• raw ingest evidence is preserved even when Claude Desktop's internal suggestion Events are intentionally excluded from the timeline
• migration v9 repairs stored Claude Desktop lifecycle state and removes previously projected internal suggestions

Validation

• python3 -m ruff check backend
• npx tsc --noEmit
• Python compilation pass for backend/app and backend/tests
• full backend suite under Python 3.12: 155 passed
• local semantic replay: 11 Claude and 20 readable Codex transcripts matched expected response, attachment, and web-search behavior; Codex gained only standalone reasoning-summary Events
• real Cursor fixture replay: plan and AskQuestion semantics matched, exactly two expected consecutive duplicates collapsed, and main/subagent interruption records flagged the retained Events
• live smoke: Codex CLI and Claude Code Desktop Sessions were captured successfully; the Claude smoke exposed and now covers the lifecycle and internal-suggestion fixes above

Remaining draft gate

• confirm a disposable Cursor live-hook Session against the current raw ingest ledger