RLCR review phase needs stagnation detection and queued-issue suppression

[Popuffu]

Problem

The RLCR review phase can enter an infinite loop: the reviewer re-scans the diff from scratch each round and repeatedly raises identical out-of-scope P2 findings. In a real humanize rlcr loop session, the implementation phase completed in 10 rounds, but the review phase ran for 38 additional rounds (48 total) — 79% of rounds produced zero output.

Root Causes

1. Reviewer has no cross-round state — Each round scans the diff independently; it doesn't know a finding was already classified as "queued" last round, so it re-raises it identically
2. No stagnation detection in the harness — max_iterations only caps the implementation phase; the review phase has no upper bound. The agent requested cancellation from Round 23 through Round 48 with no effect
3. Pre-existing code contaminates review scope — Code in the base..HEAD diff that was not introduced by the current session is treated as blocking

Suggested Improvements

Improvement	Description
Queued-issue suppression	Once both agents agree a finding is queued, record it in a session-level ledger (keyed by location + description hash). Skip it in subsequent rounds unless the code at that location changes
Stagnation detection	If the finding set is unchanged for 3+ consecutive rounds with no code changes, auto-halt the review phase
Review scope boundary	Plans should declare a change boundary. Findings outside it are advisory, not blocking
Separate review-phase max iterations	Add an independent cap for the review phase (e.g., 10 rounds). Exit with an unresolved-findings summary instead of looping forever

Impact

Observed: 10 productive rounds + 38 empty rounds. With stagnation detection, auto-halt would have triggered at Round 14, saving 34 rounds of compute and human time.