[JoyboyBrian]

Summary

Add a source-bound maintained successor to the corrected union census, with PR5001 matching capacity explicitly opt-in. The existing S=T=L/2 distinct-size gate proves its recursive prefixes simple; all other target splits bypass the predicate. Frozen engines and receipts stay unchanged.

corrected union census
  unchanged cheap tests
  optional capacity over actual functional splits
  unchanged leaf, hit, shard and report behavior

Evidence

• Before: the maintained engine did not use PR5001's capacity theorem. After: exact complete restored/accepted leaf multisets agree between predecessor, successor off/on, and an independent simple-walk oracle in five domains through n=5,L=12; 436 kernel split checks and three lost-valid-completion mutation controls pass.
• At (5,12), no-budget nodes fall from 17,360 to 16,765, while five process-inclusive timing samples give medians 4.30 ms off and 6.44 ms on. Retain opt-in status; startup-dominated tiny domains establish no speedup or larger-runtime prediction.
• Registered normal and -O replay pass. This is a scoped engine integration, not a seven-label census, new kappa, or full multiplication theorem.

Merge Danger

Door: two-way

Blast Radius: isolated

Only a new research package is added. The successor adds JSON status/counter fields; capacity stays disabled by default and is never called for repeat-capable target splits. Credit remains with Ali Romman, Nick Petalas, Mihir Srivastava, Ross Richardson, Nicholas Stone and the retained census authors. Prepared for JoyboyBrian with substantial OpenAI Codex assistance.