[CBEPX]

Research previously required Brave, and a completed research phase could leave a parent in Plan Mode waiting for another user command. This change selects native, Brave and Firecrawl search separately for each participant and uses fenced checkpoint and final events to continue analysis automatically. --manual-continue keeps the manual workflow available.

The dispatcher sends only saved workflow events to the verified native parent. The new PostToolUse and Interrupt hooks require separate trust. Cancellation, interruption, owner changes, uncertain acknowledgements and duplicate events retain their fences. Local search correlation labels remain worker-reported evidence; missing results are rejected.

Protected reviews now require stable Claude Code 2.1.321 or later. Fable 5.2 attribution has regression coverage, and expired OAuth sessions retain the authentication failure category. Model and effort defaults are unchanged.

Validation:

• 1333 unit, 144 integration and 37 E2E tests passed.
• Coverage: 93.82% lines, 85.46% branches, 98.01% functions.
• Required mutation scores: 82.65 / 74.69 / 66.39; thresholds unchanged.
• Independent source reviews have no unresolved Critical or Important findings in the final reviewed deltas.
• The v1.9.0 archive contains 97 files, all equal to the pinned candidate source.
• Real Codex 0.167.0 and Claude Code 2.1.322 research completed on this candidate: one attempt per memo and critique, no model fallback, both events consumed, all nine final read sections observed, parent completed in Plan Mode, workspace and persistent settings unchanged. Codex used native, Brave and Firecrawl; Claude used its available native and Brave routes.

Refs #27.