feat(advisor): merge final review notes without recursive reviews

[olegpulatov]

What

I improved control for final review flow.

• Merge accepted final-review findings into one age-marked delivery
• Request at most one permitted continuation for each final-review batch
• Prevent advisor-triggered continuations from recursively scheduling reviews
• Preserve default plan behavior and wake sleeping reviewers only for genuine user input

Why

Final reviews can arrive after work finishes. This keeps those findings visible and actionable without creating repeated review-continuation loops.

Depends on #12385. GitHub requires an upstream branch as the base for cross-fork PRs, so this draft targets main; review its final-flow commit after #12385.

Testing

• bun test packages/coding-agent/test/advisor/ packages/coding-agent/test/advisor-tool-call-loop-guard.test.ts packages/coding-agent/test/agent-session-advisor-suppression.test.ts packages/tui/test/advisor-config.test.ts — 323 passed
• bun test packages/coding-agent/test/agent-session-advisor-terminal-unwind.test.ts packages/coding-agent/test/advisor-context-maintenance.test.ts — 29 passed
• Exercised final-boundary delivery with merged findings and confirmed advisor-triggered continuation does not schedule another review
• bun check — passed

---

• bun check passes
• Tested locally
• CHANGELOG updated with the required attribution (if user-facing; internal issue fixes use issue links, external contributions add the PR link and contributor credit after creation)

[DarkPhilosophy]

One note from reading the final-boundary flow (no action needed if intentional):

The finally block now clears #advisorTerminalBoundaryOpen before #flushAdvisorBoundaryNotes(). With advisor.syncBacklog: off the review drain can still emit after this callback returns, so those notes land outside the merge window and deliver individually rather than as part of the merged batch. #terminalUnwindActive is still held until the next real agent start, so they cannot steer against finished work — the behavior looks safe, just not merged in that case.

Since the old comment on that finally explicitly called out the microtask gap, it may be worth restating which of the two guards now covers it, so the next reader does not assume the merge window covers every post-boundary note.

[olegpulatov]

One note from reading the final-boundary flow (no action needed if intentional):

The finally block now clears #advisorTerminalBoundaryOpen before #flushAdvisorBoundaryNotes(). With advisor.syncBacklog: off the review drain can still emit after this callback returns, so those notes land outside the merge window and deliver individually rather than as part of the merged batch. #terminalUnwindActive is still held until the next real agent start, so they cannot steer against finished work — the behavior looks safe, just not merged in that case.

Since the old comment on that finally explicitly called out the microtask gap, it may be worth restating which of the two guards now covers it, so the next reader does not assume the merge window covers every post-boundary note.

yes, thats exactly current default omp behaivour, - assistants which took too long will overfloat into later context.
just preserving it as is, to not block main agent for people who does not want it intentionally

appreciate getting another pair of eyes on the code.