[tekuuu]

PR3243's actual multiplicity-two P104 audit left the retained PR3131 nested evaluator unfinished after two 600-second attempts. This adds an opt-in exact packed representation that keeps its macro schedule, signed-word checks, inverse checks and literal cleanup executable at that boundary.

Signed Gaussian values use isolated lanes in Python integers, with checked overflow and exact floor division. Encoded words keep their declared width while storing sign extension implicitly. The original 160-bit contract, numerical order and complete protected wrapper schedule remain unchanged. All frozen inputs are authenticated before execution, and the source-specific substitutions are checked against the pinned program.

Result

• The complete 512-coordinate collector executes in a median 0.304 seconds instead of 2.565 seconds (8.44×). Full fresh-process replay, including binding and the independent oracle, improves 5.77× across three alternating pairs on the same input.
• The actual 262,144-coordinate packed nested collector completes in 285.604 seconds, followed by 4.939 seconds of independent checks. The same-input original tuple process remained unfinished at 600 seconds. No completed large-case speedup ratio is inferred from that cap.
• Every entrance, sent, returned and updated query coordinate is compared with PR3243's separately written column/parity oracle before hashing. The published predecessor receipt also matches; five sources remain unchanged and private cleanup is literal zero.
• The declared full-size plan retains 52 address gates, 104 address oracles, 432 fast-lift invocations, 1,664 inner oracle calls and both outer blocks. The expected tensor receipt was assembled from the retained output hashes, scalar gather and static call counts before the packed tensor run.

Validation and limits

python3.12 -I -B scripts/research_package.py --package research/p104-tensor-collector-execution

Complete registered CPython 3.12.14 replay passed at executable commit b9a2c1efdfd862aab7cd4cfd53529a81e0c76f00. The read-only Git-free export passes eight groups and all small differential cases under optimized CPython 3.14.7. All 37 publication package/input byte checks, metadata/pins and whitespace checks pass. Final amendments change only documentation and timing evidence; executable and deterministic result bytes are unchanged. Hosted CI is separate.

Eight behavioral groups cover all 2,401 signed local pairs, lane isolation, negative floor division, 257-bit inputs, full wide macros, nonconsecutive axes, encoded-width mismatches, overflow and inexact shifts. Original unguarded-CNOT, omitted-phase and omitted-cleanup controls are retained. Corrupted inputs are rejected before vendor execution. Source bindings, deterministic receipts, timing evidence, the representation proof and open questions are in the package.

This is verification tooling and finite implementation evidence. It does not execute or benchmark the fast physical C backend, alter the recursive invoice, admit a supplier or claim a new kappa. The unchanged P104 full consumer still has the 599/584 zero-saving obstruction. The original tuple evaluator's capped large run remains incomplete; no original full-size completion is claimed.

Only one scoped package is added. Base is 463113d675af7410ac91d53a4fed51ec6494eb9a; PR3131 is pinned at 0c9e344d5a5295a218558fa4a51d867f39ff5311, its PR3109 archive at 15eac811648182bdac7254a91fd9b841b1414ebc, and PR3243 at 993e2d4ba993d9818c6d7c5ae1e35980d1ab3ebb. Pending inputs are retained verbatim, so canonical CI needs no foreign Git objects. Selected and frozen predecessor files are unchanged.

Credit Jose Manuel/PR3131/3109 for the program; commdata2338/PR3243 for the actual fixture, independent oracle and released comparison handoff; Tekalign/PR3025/3033, Ross Richardson/PR3072, Douglas Colkitt, ywz65, Vasek/flyingviper and the retained native/compiler/physical lineage. PR3134's wrapper contraction is adjacent and is not imported. New work by Tekalign (@tekuuu), with substantial OpenAI Codex assistance in design, implementation, proof, checking, measurements and writing. Apache-2.0; no independent human or formal-kernel review claimed.

Canonical investigation.