RLCR: Reduce whack-a-mole fix cycles with subsystem audits and velocity tracking

[lczong]

Context

During a 7-round RLCR session (continuation of a prior 18-round session), rounds 0-2 were highly productive, but rounds 3-6 exhibited a "whack-a-mole" pattern: four consecutive rounds each fixed a single localized issue in the same subsystem, where each fix revealed the next related defect. These 4 rounds consumed 57% of implementation rounds but contributed ~15% of total value.

Observations

1. Narrow fixing without subsystem audit: After a review identifies a defect in a specific area, the implementer fixes only the literal finding without proactively auditing surrounding logic. This leads to serial single-issue rounds.

2. Reviewer identifies issues one at a time: The reviewer found genuine defects in every round (zero false positives), but surfaced them one per round in the same feature area, when a holistic design review could have caught most at once.

3. Binary stalled/advanced verdict masks inefficiency: Each round technically "advanced" but progress velocity was very low. The methodology has no signal for "advancing inefficiently."

4. Reactive fix cycles for design-level issues: When a feature needs design-level refinement (not just bug fixes), the implement-review-fix cycle is suboptimal.

5. Continuation sessions inherit stale assumptions: A round-0 review immediately found fundamental issues the prior 18-round session had missed.

6. Learning mechanism underused: The implementer marked "no lessons" across all 7 rounds, even during the 4-round iterative refinement cycle that was itself a lesson.

Suggested Improvements

#	Suggestion	Mechanism
1	After review feedback in a specific area, require self-audit of the entire affected subsystem	Add to round contract template: "If prior review found issues in area X, list the invariants you verified for the full subsystem"
2	Reviewer should evaluate the full design contract of a newly introduced feature	Add to reviewer prompt: "If this is a new feature, evaluate all its documented invariants end-to-end, not just the first defect"
3	Add a "velocity" signal alongside stalled/advanced	Flag when 3+ consecutive rounds each fix a single localized issue in the same area; prompt a comprehensive subsystem audit
4	For design-level issues, require a design doc before implementation	When a review identifies a design-level issue (vs. a simple bug), the next round contract should include an invariant list
5	Trigger comprehensive re-review on session resumption	When an RLCR session resumes after interruption, the first review should re-evaluate foundational assumptions
6	Treat the learning system as a process-improvement tool	Prompt the implementer to record process patterns (not just technical novelties) in the learning mechanism