Give Reflection an evidence-led learning loop

[hamzamerzic]

Summary

• give Reflection bounded receipts for its own work and the chat-agent work it reviewed
• turn repeated tool friction into evidence-backed leads rather than hard-coded diagnoses
• record interview outcomes, experiment hypotheses, observations, and follow-ups in compact private ledgers
• hand Memory recall health and hindsight evidence into Reflection without copying private chat content
• keep reports responsible for useful conclusions while leaving ordinary settings focused on schedule and optional model overrides
• make overlapping scheduled runs fail before they can clear an active run's evidence
• fail closed when expected interview coverage cannot be established, and keep unreported provider cost unknown rather than zero

Why

Reflection needs enough scaffolding to work like a thoughtful maintainer: inspect recurring failures, interview the agents that did the work, choose the smallest owning fix, record a hypothesis, and revisit the result later. The ledgers are evidence and continuity for intelligent judgment—not a metric optimizer or a new dashboard.

This is a distinct follow-up to #46. That change correctly removed effort controls and persistent run telemetry from the user interface; this one builds the private learning machinery behind the simpler surface and preserves provider-default effort.

Validation

• npm test — 41 JavaScript tests and 100 Python tests passed, including shell and Python syntax checks
• validated the merged result against the current upstream base
• rendered the simplified Settings surface at the owner viewport

[hamzamerzic]

Review completed against exact merged head d6884f72376ceba6dcefac1bb4db640a054b0bd7.

I reviewed the unattended-run lifecycle, overlap lock placement, evidence staging and freshness manifest, experiment/learning ledgers, question handoff, privacy-safe tool-friction aggregation, model-usage receipts, and report contract. The no-overlap lock is acquired before mutable run inputs are cleared, the evidence bundle fails closed on stale sources, questions remain declarative/non-blocking, and retained diagnostics avoid raw command/transcript content.

One blocker was repaired before merge: missing provider cost was being coerced to zero in the friction report, which mislabeled unreported cost as free. Cost totals and per-run values now remain explicitly unknown unless a provider reported them, with a reported-run count and regression coverage. Full app suite: 41 Node tests + 101 Python tests passing, plus shell/Python syntax gates.