Prove rounded trace transfer for QQ16 streamed recovery

[Clear1nsight]

PR #4015 changes private query storage into persistent hosted output banks and interleaves query additions with native calls. This contribution checks whether that rescheduling creates a separate rounded-arithmetic obligation.

The complete #3985 and #4015 words have exactly the same 3,392 value-producing operation dependencies and occurrence multiplicities: 2,816 ordered scalar ADDs, 352 cover maps and 224 signed native calls. All 224 output roots match; the inverse-B1 checkpoint matches at old2852/new2730; the native input/recipe sequence also matches in order. The checker uses exact structural tuples, with no reassociation, inverse cancellation or floating-point sampling as its proof criterion. RESULT.json retains the complete event bijection.

This proves relative bit/error transfer under an explicit finite-primitive contract: paired operations must use identical widths, binary scales, rounding/overflow conventions, scalar product/add order and native interior recipes; persistent stores preserve bits; CLEAR gives a common exact zero; native workspace is disjoint and reset; kernels have no hidden interleaving-dependent state. A valid finite-format range/error certificate for #3985 then transfers to #4015 without a new rescheduling error term. It does not remove arithmetic error sites or provide the missing certificate itself.

The source-bound adverse cases make the premises testable. At old2140/2141 versus new1716/1764, query values3,1 fit a four-bit old private accumulator but overflow a three-bit new owned bank. At the literal -1/4 response old2563/new2537, signed floor and truncation disagree. A deliberately stateful native countermodel already distinguishes old9/new7 through the changed number of earlier scalar ADDs. These are explicit boundary/countermodels, not alleged failures or reachable full-producer trajectories of the published kernels.

The authenticated #3303 proof and retained #3652 proof certify the literal F0/B1 producer, with quantum h*2^(-71-19*(lambda-1)) and width P+215+38*(lambda-1). They explicitly leave F17 consumption outside their scope. Neither supplies the decoder-interior formats or proves that old private F accumulators and new persistent Q banks have the required common arithmetic. #2965 still requires the implemented decoder's epsD; its older #2736/#2707 invocation is not automatically a certificate for the changed #3985 rectangular word. The useful next step is to certify one actual paired decoder implementation on the shared DAG, then bind caller ownership and physical access.

Only research/qq16-rounded-trace/ is added. Two canonical scope sources are declared for invalidation; the two pending WORDs and #3652 proof are narrowly retained with exact heads/blobs/hashes. No predecessor executable or unmerged Git history is required. Original native/frame semantics remain inherited; no full physical execution, supplier, kappa or frontier claim is made.

Validation on CPython3.12.3: registered primary and all six boundary groups pass, including normal/optimized read-only, normal/optimized Git-free read-only, canonical/retained source authentication before parsing, full mapping and typed receipt corruption, and overwrite refusal. Nine semantic controls cover initialization, recycled storage, query order, output restoration, native/private overlap, signs and borrowed ownership. Signed-format controls use an exact-rational/two's-complement oracle. The coordinating agent and a separate physical-transition agent each reviewed the proof/code and independently ran optimized warning-strict primary replay successfully; scope is recorded in REVIEWS.json, not represented as human or proof-kernel review.

Publication head: aeeca9ffb50203b56924ee3f1325deab4029f4d9, created through GitHub from canonical 20ecf2891d80b5acf18d980b2164fb44ea7faa90. Exact-byte/source-history preflight passes for all16 package/input paths; staged and committed whitespace checks pass. The official committed export and hosted CI are separate statuses below.

Assigned parallel round in the canonical investigation. Main/priority6/selected record were refreshed unchanged at11:15UTC; binding baselines passed. Targeted live and local overlap was partial; no exhaustive novelty claim. #4030 changes a distinct normalization/frame question and is not imported.

Credit Chris R/Clear1nsight #3985/#4015; ywz65 #2707; Arham Tahir #2681; Ali Romman and retained QQ16/native/geometry lineage; Nick Petalas #3303/#2965 and the precision predecessors; ToikY #3652. Substantial OpenAI Codex assistance is disclosed. Apache-2.0.

The official committed16-file export and offline integrity check also passed at this exact head. Hosted CI remains separately observable.

Hosted verification addendum for PR #4043

PR: #4043

Published head: aeeca9ffb50203b56924ee3f1325deab4029f4d9.
Canonical base: 20ecf2891d80b5acf18d980b2164fb44ea7faa90.
Hosted synthetic merge: f74f5eb225d88f70271d925d48e523346247177f.

The Registered research packages run 39243839930 completed successfully on 2026-10-19. Both completed job logs were read and retained. Both checked out the exact synthetic merge of the published head into the canonical base, using CPython 3.12.15. The PR was re-read after completion: open, unmerged, head/base unchanged, maintainer modification disabled. The local publication worktree is clean.

The plan job 118030830286 ran 265 onboarding helper tests in 58.596 seconds, with 11 native-Windows-only skips. It selected exactly research/qq16-rounded-trace.

The replay job 118031606797 reported PASS_QQ16_ROUNDED_TRACE_EQUIVALENCE, the complete 3,392-event mapping, equal native-call sequence, 3,564 unique structural nodes, and inverse-B1 checkpoint pair (2852, 2730). All six registered boundary groups passed in 11.827 seconds, including normal/optimized read-only and Git-free modes, retained and canonical source authentication, full mapping and typed receipt controls, and overwrite refusal. The final result was PASS_REGISTERED_RESEARCH_PACKAGE research/qq16-rounded-trace outcome=partial. The following committed-inputs-unchanged check succeeded.

The earlier publication preflight and official 16-file offline export both passed. Archive SHA-256: 3b93f71783ef5d8c95e8bf55c04a60ca65b1232f395a80e45eeb7f368daa72b5.

Retained full plan log SHA-256: 1bebc83f30a316971d601b12b2ac85f354d6536cf6a89df91bea44be81c9b7d2.
Retained full replay log SHA-256: 561c73e8ecae5ec3c12f8ee297798b277a0f8cf477ccbe3799058d5da87d0791.

This establishes reproducibility of the conditional rounded trace transfer result. It does not supply the missing common decoder format/range/error certificate, a physical tape implementation, or a qualified multiplication constant. No scientific files changed after publication.