Resume Reflection from its last completed run

[hamzamerzic]

Summary

• resume evidence collection from the last successfully completed Reflection run
• keep failed, dry, quota-blocked, and partial runs from advancing the checkpoint
• stage outstanding chats, feedback, answers, activity, tool friction, and recall evidence oldest-first
• migrate once from existing run metrics and cover restart, pagination, ordering, and empty-input cases

Why

Reflection previously gathered fixed recent windows and newest-item samples. If a scheduled run was missed or failed, older evidence could fall outside the next run's window and never be reviewed. A successful report could also coexist with a partial input gather without a durable boundary saying what remained.

This change introduces one completion checkpoint owned by the run. Backlog-producing inputs all resume from that boundary, and the checkpoint moves only after a real successful run writes a fresh brief with complete inputs. A missing question-answer directory is a valid empty queue; transient source failures remain retryable instead of being skipped.

PR #38 introduced useful recall-quiet detection. PR #42 proposes extending that bounded baseline. This takes a different route: it removes the sliding automatic verdict and gives Reflection ordered recall activity since its last completed run to compare with actual chat opportunities.

Tests

• npm test (41 Node tests, 77 Python tests, shell syntax and Python compilation checks)
• app validator (0 errors, 0 warnings)

[hamzamerzic]

Möbius agent review

Result: No blocking issue found in the current revision. This is the preferred Reflection fix: the full diff resumes from a durable completion checkpoint, keeps recall evidence ordered, and replaces rather than layers over the sliding baseline in #42. Close #42 before landing this one.

I reread the full diff and checked the current GitHub workflow result; the reviewed revision is green unless the queue note above calls for a later rebase.

This is a disclosed agent review posted by the PR author’s account, not an independent maintainer approval.

[hamzamerzic]

Expanded-scope follow-up (no blocker). I traced the checkpoint from migration and atomic advancement through activity/chat/question staging, deleted-chat pagination, oldest-first ordering, manifest freshness/hash validation, runner completion, partial-input behavior, and the health/friction consumers. The checkpoint advances only after a real successful run writes a fresh brief, while partial or retained inputs are labeled rather than silently treated as current. The broader test suite passes (77 Python tests plus the fetch staging test). I found no blocker or redundant resume mechanism.