Claude-native telemetry: delivery correlation (shipped-vs-reverted yield, verify loops)

Problem

Telemetry says how much was spent, not whether it shipped. The highest-value signal ties Claude activity to delivery outcomes: kept, reverted, or abandoned, and how many edit→verify cycles correctness took.

Capability

Correlate normalized telemetry with git history to classify productive vs reverted vs abandoned work, and count verification loops.

Behaviour

• Yield classification: productive (kept to tip), reverted (e.g. This reverts commit <sha>), abandoned (unmerged branch); attribute window spend to each class.
• Edit → verify → edit retry counter within a session (edit, then test/build/lint, then re-edit same target).
• Rework signal via repeated rewrites of the same file (incl. per-session file-history snapshots).

Acceptance criteria

• A fixture repo with a revert classifies as reverted; its window spend counts against reverted yield.
• edit→test→edit increments the retry counter; edit→passing-test (no re-edit) does not.
• Findings cite commit SHA + session/turn evidence.

Dimensions: Delivery Predictability, Test Loop Completeness, Quality & Maintainability. Tier 3. Depends on #36 + existing git analyzer. PRD: docs/prd/features/05-delivery-correlation.md.

Start with docs/prd/features/05-delivery-correlation.md, dependency #36, and the existing git analyzer. Use a fixture repo with a revert and edit→test→edit history to verify yield classification, retry counting, and rework detection; done means findings cite commit SHA plus session/turn evidence and spend is attributed to the yield class.