[ywz65]

PR2195 leaves the work allocation and routing bill for a complete BIT invocation open. This package gives a conditional node compiler: gather all disjoint low-owner packets once, align them, run the whole local word on the private copies, then inverse-align and scatter. It charges two full scans per touched family, two owner-route actions per role, work copying and cleanup, and every local address child.

The selected PR589 census has 288 calls and 661 roles per call. The resulting schedule has 1,322 nominal owner routes per call and 213,056 family sweeps across the 288 calls for one low class. The node reserve is 663*L + W_max, conditional on completed route/callee work contracts. A separate family layout is an explicit entry premise; splitting a concatenated live array costs an additional linear pass and copied storage. These are a parametric bill and a node reserve, not a complete recursive fixed-tape allocation or numerical supplier admission.

Evidence in research/bit-whole-packet-staging/:

• Independently checks 181,056 selected helper support pairs and all 661 Boolean columns in both 27,234-event local words.
• Sequential GL2(Z/9) gather/route/word/scatter implementation covers all 3,888 matrices, three unused spectator values and 17 payload bits. A separate full-ring oracle checks all 11,664 symbolic record columns, including two calls that reuse changed packets.
• Seven semantic controls reject owner aliasing, missing carries, wrong inverse routes, omitted restoration, incomplete spectators and stale state. Peak finite-fixture scratch is 20,655 bits and all private work is cleared.
• Git-free relocated replay is read-only and rejects real source/receipt corruption. A separate AI peer reviewed the proof/code boundaries and independently passed the optimized replay.

python3.12 scripts/research_package.py --package research/bit-whole-packet-staging
python3.12 -I -O -B research/bit-whole-packet-staging/verify.py

Registered replay passed locally on Python 3.12.14. The data-movement meter excludes Python controller/allocation costs; the proof charges counters, preparation and head returns symbolically. Complete weighted callees, global recursive parking, good/fallback integration and the numerical routing invoice remain open. No selected frontier, new kappa, cold selected replay, hosted CI pass, independent human review or formal-kernel proof is claimed.

Credit Jose Manuel / jmbarrancoml's PR2195 for the complete-packet separation, per-ADD selector and row layout; alephnan's PR1627 for the whole-call/frame proof; eumemic's PR490/555, DreamingOfClouds' PR589 and the retained owner/stream lineage. PR2437's ideal changed-word binding is compared explicitly; this contribution does not claim its full physical admission. Exact source bytes and external heads are pinned. The scope change from the historical H2/K card and bounded live overlap are recorded in TASK.json.

Prepared for ywz65 with substantial OpenAI Codex assistance in proof, code, testing and writing, including separate AI peer review. New work is Apache-2.0. Canonical main at final compatibility check: 0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2, priority revision 4.

Final overlap refreshed 669 open PRs at 11:51:45 UTC and 1,192 Discussion records at 11:55:21 UTC on October 15, 2026, with complete pagination and no fetch warnings. Closed work and code-diff coverage remain bounded as stated in TASK.json.