fix(review): require explicit follow-up delivery

[limityan]

Summary

• make review subagents completion dependencies by default so the owning agent can return one merged review result
• allow later review delivery only when the user explicitly requested not to wait, via run_in_background=true plus allow_review_follow_up=true
• classify built-in and custom reviewers through the workspace-aware registry, including cold reused sessions, without changing non-review background tasks
• fail closed before fully restoring cold review sessions, while using a zero-turn read-only preflight to avoid duplicate transcript loading

Product behavior

Scenario	Behavior
ordinary review request	waits for review results; parallel foreground reviewers can still run concurrently
explicit "background, do not wait" review	returns now and delivers the review in a later follow-up
non-review background task	unchanged
managed Review / DeepReview flows	unchanged

Scope

This change does not add natural-language review classification, generic task groups, UI changes, result deduplication, or new cancellation semantics. allow_review_follow_up permits later delivery but does not detach the task from normal parent cancellation.

Validation

• cargo +nightly test -p bitfun-core agentic::tools::implementations::task::tests --features product-full -- --nocapture (68 passed)
• cargo +nightly test -p bitfun-core review_follow_up --features product-full -- --nocapture (7 passed)
• cargo +nightly test -p bitfun-core cold_ --features product-full -- --nocapture (6 passed)
• cargo +nightly test -p bitfun-core agentic::agents::registry::tests --features product-full -- --nocapture (23 passed)
• cargo +nightly check -p bitfun-core --features product-full
• node scripts/check-core-boundaries.mjs
• pnpm run check:repo-hygiene
• rustfmt +nightly --check --edition 2021 ...
• git diff --check gcwing/main..HEAD

Independent review

Independent architecture and product-design reviews found and closed issues around misleading detach semantics, cold-session restore side effects, workspace-aware reviewer identity, missing success-path coverage, user-facing recovery guidance, and duplicate transcript reads. Both final reviews report no remaining actionable findings.