[RetroGameDeveloper]

What changes

• After a random rewrite improves the saved candidate, deep search immediately tries up to 32 local proposals against its new diff, chaining saved improvements. All work shares the original deadline, compile allowance, cache and cancellation/claim handling.
• Add bounded constant/copy reuse and final integer shift/scale reuse candidates to shared automatic/manual repair paths. The normal compiler and exact judge remain authoritative.
• Reverify score-zero/open functions and record a fresh exact verdict even when the scalar score cannot improve below zero.
• Recover an otherwise unbuildable ROM draft when one unambiguous plain shared return declaration conflicts with its definition. Try a candidate-only return alignment, preserve the broken source in attempt history, guard publication against changed inputs, and rebase its measured score honestly. The shared header is not edited. This may lower the displayed draft percentage.
• Retry the two formerly blocked pass states under the new rules, preserving unrelated completed passes. Explicit Repeat remains available for older completed functions.
• Retain local-repair family evidence and refinement burst counts for future analysis.

Validation

• 101 focused tests pass on the final commit, including the random/local feedback loop in ROM and imported-project modes, shared limits and leases, baseline recovery/races, exact-verdict bookkeeping, and a pinned real-MIPS compiler test recovering an exact synthetic object through constant reuse.
• Full local suite: 2114 passed, 156 skipped (514.88 seconds). The final winner-evidence follow-up was also verified with the focused suite. Ruff, whitespace checks and the GUI regression script pass. Platform CI is still running.
• The macOS package builds and the installed app restarts successfully. All 295 packaged source/UI files match this checkout; the live project-wide pass resumes automatically.
• Private five-function regional racing-game replay: 401 candidate evaluations with a 100-candidate cap per function. The already-exact/incorrectly-open function becomes matched after one fresh judge call. The conflicting draft now compiles (fresh baseline 215) and improves to 200; its old unbuildable draft had a stale score of 115. Three register-only near misses retain scores 5, 10 and 15. No new source match is claimed. No private game source/data is included in this PR.

Limits

This implements the prioritized feedback/register work and two recovery fixes. Adaptive project budgets and a pool of alternative search candidates remain planned. Return recovery only handles plain scalar declarations with unchanged compatible parameters; ambiguous declarations stay with review. Existing exact-match rules are unchanged.