decomposer(feedback): execution outcomes back into the drafter

[JSONbored]

Problem

Every executed issue is a labelled example — this phrasing, this sizing, this acceptance
detail produced this outcome — and the loop currently does not close. There is a post-merge
feedback path into the drafter, but nothing systematically feeds the quality score back.

Scope

• Feed per-issue quality scores and their component signals back as drafting signal:
  which phrasings, sizings, and acceptance-detail patterns correlate with clean first-pass
  merges
• Persist the corpus of (issue text → outcome) pairs in a form usable for evaluation and
  future training, with the same held-out-split discipline used elsewhere in the repo so
  improvements are measured rather than asserted
• Report drafter performance over time against Epic: issue-quality measurement — score a scoped issue by how cleanly it executed #10076's score, per repo and per drafting
  method, so a regression is visible

Acceptance

Drafter changes are evaluated against held-out historical outcomes rather than judgment,
and the trend is visible.

Note

This is the compounding piece — decomposition quality improves from real execution rather
than synthetic data, and the corpus grows with every merged PR.

[linear-code]

JSO-7092