Verify R2 native cleanup and all 91 private zero returns

[24kHandsome1201]

PR3202 left the complete R2 cleanup's native semantics open. This focused package checks the exact PR3044 base/extra providers and proves that all 91 clean private arrays return to zero at their actual frees for arbitrary live entrance arrays.

The new checker binds all 5,208 cleanup gates and all 5,585 phase events, including private namespaces, allocations/frees and residual post-adapter placement. The leading 204 gates are independently matched to R2's reversed restore inverse. Each of the 72 used cleanup warp recipes per provider is checked on all 512 complex columns, including global phase, in both orientations. Whole-record/source binding distinguishes four provider differences: 76 distinct native records are verified.

A separate exact rational free-group interpreter checks all 632 arbitrary live entrance arrays. All 91 private rows are zero immediately before free and all 112 target rows are unchanged. The proof composes the checked native warp identities with scalar gates through the common FULL coordinate change. Private zero-return needs no reachable-state premise from the earlier prefix.

The producer/compiler and original cleanup evaluator are never imported or executed. Native arithmetic reuses Sagar Thakrar's pinned independent Gaussian-integer kernel. The free-group cleanup method credits commdata2338's earlier PR2269 buffer-pooling work. Five compressed files retain six authenticated native artifacts from PR3202's fresh generation; they were reused, not regenerated in this session, and are checked against PR3044's published receipt.

Validation:

• Registered Python 3.12 replay passed in 18.816 seconds.
• Isolated optimized replay passed in 12.436 seconds with identical receipt.
• All 52 semantic mutations reject after authentication by named predicates, including native phase faults, coordinated inverse corruption, private aliasing, wrong closing writes, dirty entrances and nonzero early frees.
• Metadata/source pins, whitespace checks and all 21 package/input publication-byte comparisons pass.
• Published head: 87d63a82defff1ffeea173743106fb97903cbbd6. Hosted CI is separate.

This is a finite partial audit, with no new kappa, storage saving or supplier admission. Rational scalar execution and owned clean allocation remain interface premises. Complete native-prefix/live restoration, actual executor/owners, precision/tape, all-size/compiler and supplier/outer obligations remain open. The 429 forward and 397 dual changed live rows show why cleanup alone is not the complete invocation identity.

Task independent-moment, scope refined to PR3202's explicit native-cleanup handoff. Scientific source: PR3044 1f922b25abff2d7fd87fa5b29619160862a06b59. Canonical main remained 463113d675af7410ac91d53a4fed51ec6494eb9a, priority revision 6; selected frontier unchanged. Exact card, source pins and bounded overlap coverage are in TASK/SOURCES.

Overlap: ToikY's PR3038 V45 cone receipts cover 30 FULL cleanup warp classes forward and 8 backward per provider on PR2979; this checks all 72 each way on R2. commdata2338's concurrent PR2979 peer-private-free/dependency-cone audit is credited. Tekalign's PR3202 kernel-acceleration claim is complementary. No priority is claimed for the general methods.

Proof, scope and reproduction · Canonical investigation

python3 -B scripts/research_package.py --package research/pr3044-native-cleanup

Next: complete the actual R2 native prefix and compose its live endpoint with this cleanup; independently bind a real private allocation/free interval to its physical owner and full movement/precision bill. README/TASK give the first steps and blockers.

Credit Nick Petalas, Jeremy Kothe, Marius-Paul Dumitrean, Hiroki Kumazaki, Sagar Thakrar, ToikY, commdata2338, Cinna Davis and the retained native/geometry/owner lineage named in the README. 24kHandsome1201 with substantial OpenAI Codex assistance; solo agent, no independent human or proof-kernel review claimed. Apache-2.0.