Audit exact total tape traffic for high-field doubling and halving

[petalas]

Summary

research/high-scaling-total-traffic/
  PROOF.md                 exact all-parameter local traffic invoice
  traffic.py               independent closed formula
  source.py + observe.py   authenticated literal execution and origin tracking
  verify.py + tests        complete receipts and rejection controls

Independently audit PR2915's kernels.doubled. For odd Q>=3 and positive canonical spectator extents B,S, doubling and halving have the same exact logical tape traffic, including Count setup, every countdown, copies, rewinds and cleanup. The written source-level bound is T<1024 BQS. The exact construction count supplies a separately charged factory interface T+aA; allocator work is not assumed free.

Credit Jose Manuel/jmbarrancoml for the algorithm/caller, David Arias for the all-size caller, and Arham Tahir/PR2877 for the shared countdown identity. Their source lineage and exact heads are retained. This adds the independent scaling composition and checker, not another algorithm or kappa.

Base 3aa52a3b2f8c08994978f97d98c62cfa8e4916fb, priority revision 5. All fourteen provenance/input rows match refreshed main 389fa10d714f5b40afe624908e487eb5e424498d; the patch applies there. Canonical claim and handoff.

Evidence

• Before, PR2915 provides the implementation, finite endpoints and a linear-cost argument. After, this separately derived exact invoice matches 390 complete literal calls and all 152,352 payload origins in both directions, plus 45 shared-countdown cases.
• An endpoint-preserving extra-countdown mutant raises 10,354 operations to 49,874. The exact invoice rejects its 39,520 excess. Wrong-direction, omitted-cleanup, malformed-input, source/receipt and canonical-width controls also pass.
• python3.12 -I -Werror -B scripts/research_package.py --package research/high-scaling-total-traffic passes all three registered commands, including optimized acceptance and thirteen behavioral groups. Optimized CPython3.13.5 replay/tests also pass on Linux/WSL2.
• All 144 research-workflow tests pass without skips. Strict lint/format, exact staged diff and the 23-file committed-byte preflight pass.

Successful local source-slice evidence only. Physical allocation/placement, complete routes/caller, failure-path cleanup, supplier admission, human/formal review and hosted CI are not established. Selected result and frozen inputs are unchanged. Substantial automated assistance; author-checked evidence.

Merge Danger

Door: two-way. Remove the isolated package to revert.

Blast Radius: research. The opt-in replay adds no production or selected-frontier changes.