Context. ADR-0026 added a net-diff floor to fix the pilot-era ~100-line over-decomposition problem. It worked and overshot: the consumer pilot's merged sdd-PRs ran median 694 net lines (target ≈400), range 79–1765, with three PRs >1300. Per-PR review forensics: the oversized PRs carried the heaviest convergence loops — the 1504-line PR drew 66 review comments across many passes; the >1300 PRs concentrated the review-loop cost that dominated the run's back half. No ceiling exists.
Problem. The floor has no counterpart. Nothing splits a task whose materialized diff blows past the target band, and the biggest PRs are where the review loop (the run's throughput bottleneck) costs the most.
Fix:
- A sizing ceiling (soft band ~400, hard cap TBD) paired with the ADR-0026 floor.
- Split-at-PR-time: when a task's diff exceeds the band, sdd-execute splits along commit/scope boundaries into stacked PRs rather than shipping one oversized PR.
- Advisory at plan time (triage flags likely-oversize tasks); enforced at PR time.
Acceptance criteria:
- A materialized PR over the hard cap is split (or blocked with a split directive), not merged as-is.
- Triage flags tasks whose estimated diff exceeds the band before dispatch.
- Sizing distribution (median / max / over-band count) reported per feature.
Relates to: #272 (the unit-count lever — co-design so the ceiling doesn't re-inflate sub-issue count; this is the PR-diff lever), ADR-0026 (the floor).
Context. ADR-0026 added a net-diff floor to fix the pilot-era ~100-line over-decomposition problem. It worked and overshot: the consumer pilot's merged sdd-PRs ran median 694 net lines (target ≈400), range 79–1765, with three PRs >1300. Per-PR review forensics: the oversized PRs carried the heaviest convergence loops — the 1504-line PR drew 66 review comments across many passes; the >1300 PRs concentrated the review-loop cost that dominated the run's back half. No ceiling exists.
Problem. The floor has no counterpart. Nothing splits a task whose materialized diff blows past the target band, and the biggest PRs are where the review loop (the run's throughput bottleneck) costs the most.
Fix:
Acceptance criteria:
Relates to: #272 (the unit-count lever — co-design so the ceiling doesn't re-inflate sub-issue count; this is the PR-diff lever), ADR-0026 (the floor).