Prove QQ16 final-only loss despite 64 initial query coordinates

[alephnan]

The 64 bank-local initial QQ16 query coordinates do not replace the selected state recorded after B1 in PR #2174. On the actual regular owner cover, this package proves that the specified final-only observation map loses at least ceil(8|G|/7) dimensions of pair 0's original selected input, even with final selected fields, completed-X outputs, and all forward histories granted free.

The scope is the retained 19-event local schedule with ordinary-only bridges/terminal and no added selected intervention before F2. It does not cover the separate first-six-selected-swap controller or all possible sets of 64 queries. No supplier, native-cost, precision, or improved kappa claim is made; the selected result stays unchanged.

The proof uses the actual-route two-moment ranks to show that two constrained query images intersect after the required ordered selected transport. Their rational sections yield a nonzero family invisible at every allowed final observation. This upgrades philippeVV's pinned four-bank finite-model failure; the prior finite failure, record choice, and moment census are credited explicitly. Other direct predecessors are Ali Romman #1969, Ross Richardson #2104, Adam Gall #2093, and @commdata2338 #2174; related higher-moment work is also credited.

Review research/qq16-final-record-obstruction/PROOF.md for the theorem and README.md for the contract, source lineage, and open next questions. The package contains exact source pins, independent forward/backward centre reconstructions, full owner-route comparison, a 19-event symbolic trajectory, separately factored rational algebra fixtures, and a complete read-only result receipt.

Validation passed:

• python -I -B scripts/research_package.py --package research/qq16-final-record-obstruction (normal and optimized replay, independent controls, relocated Git-free replay, and tampering rejection).
• Selected-source baseline bindings, unchanged, using an extended-length Windows path.
• Two separate agent reviews found no mathematical/source-binding blocker; this is not human or formal proof review.

The selected full cold replay was not run because no selected supplier or source changes. Compact witness support, bitlength, precision, and native implementation remain open. Main and fork were refreshed at submission; detailed coverage and timestamps are in TASK.json.

Canonical claim and handoff. Task uniform-transfer-audit, refined to this explicit final-retention handoff. Substantial OpenAI Codex assistance in proof, implementation, writing, and separate agent checks; new work is Apache-2.0.

[philippeVV]

Pointer: your open question (OQ2, full recovery from final records with retained d_A) is addressed in #2074 (comment). In finite noncommuting models, d_A repairs bank 0 only. Banks 1 and 2 keep losses that no pair-0 record can repair. INIT64 + d_A on banks 0–2 (88 arrays) suffices. Whether it pays depends on how a retained array is priced per call. This is evidence, not an actual-cover proof; artifact and checkers are linked there. (Prepared with Claude Code assistance.)