[JoyboyBrian]

The public finite compiler retains fresh Residual wrappers for repeated source keys. This scoped package answers the sharing handoff from #3391 and #3650 (selected-compiler-source-bridge): an opt-in post-compilation pass shares equal definitions within a mode, with a conditional simulation proof for the unchanged physical caller.

The pass checks complete residual segments, exact segment-owner identity and every definition field, including code and plan. Its contract requires deeply unchanged fields and read-only, identity-insensitive observers. The original geometry identity guard and source callbacks remain in place. Malformed segments, same-key/different-plan definitions and cross-mode collisions reject; controls also demonstrate why identity observers and stale geometry are outside the proof.

Validation:

• Complete normal GL2 before/after runs: 24 retained wrappers become 12, with all 963 operations, field-based traces, physical counters, output and erasure preserved.
• Unchanged nested fixture before/after: 12 child calls/returns, 156 residuals and all 9,216 payload positions, including codec/parking receipts. Both normal runs complete in approximately 117 seconds each. All nine rejecting controls additionally pass in normal mode.
• Optimized replay of the same complete checks from a committed Git-free export passes: 257.817s wall, 249.989s user CPU, 1.503s system CPU, 81,821,696-byte peak RSS. Nested before/after runs take 126.44s and 123.06s. These are replay resources, not an optimization benchmark.
• Package metadata/source pins, committed-byte/source-history preflight and 100-file export integrity pass. Python 3.12.2, standard library, macOS arm64.

This reduces wrapper identities retained by returned Code after old references are released. It does not remove original emitter allocation, setup peaks, operation tuples, direct references, LOW enumeration or full compiler cost. The earlier 32-visits-per-key coefficient remains conditional on its source-census hypotheses and equal definition fields; selected GL64 is not expanded here. No stock, supplier admission, selected frontier or κ changes are claimed. Historical loader admission is explicitly outside the reduced runtime experiment.

Canonical claim: #2319 (comment)

Credits and open next questions are in the package README. Prepared by JoyboyBrian with substantial OpenAI Codex assistance; no independent human or formal acceptance is claimed.

Submission checkpoint: canonical main was fetched again and remains b58ad9ef4d4e861b0eaa5dba41742bdc072dca51, priority revision 7. The refreshed complete inventory (2026-10-22 19:10:08 UTC) covers 512 open PRs, 3,673 Discussion records and 5,860 base local notes. Scoped search and live predecessor heads were rechecked; no competing result was identified in this stated coverage. Closed PR diffs, arbitrary nested code and unpublished work remain outside it.

Hosted replay budget correction: Ubuntu 24.04 / CPython 3.12.15 completed the first nested execution in 274.67s, then hit the former 360s command limit while running the second execution (failed job). Both registered commands now have the permitted 600s limit, totaling 1200s. Both full normal/optimized runs, scientific code and RESULT.json remain unchanged. Metadata/source-history/committed-byte/export checks pass; hosted success still requires the updated CI run.