[codyw912]

CRT scanner-owned-mask consumption/disposal bridge (partial)

Result

Adds the narrow #4208/#4201 scanner-owned-mask consumption/disposal interface to the inherited #2250/#4328 sparse repair: marker-driven unit-action split, full-key stable radix reinsertion, physical clearing and head-return gates. It consumes the scanner's actual mask in place, rather than rebuilding it from a host good list.

The existing #4169/#4100 fixture is exercised as ten separate repaired-layer invocations, not an uninterrupted all-tape caller: five forward Beneš layers and five inverse layers over 4,096 labeled payloads, with 29,440 hole-key calls. #4169 already supplies current-state inverse-word keys and this exact fixture/endpoint. Label 780 reaches 777 and returns to 780; all 4,096 payloads restore.

Host Python key arithmetic is uncharged and not a tape evaluator. Macro/raw streams and the inter-layer payload handoff remain host-built; blank private-role reuse does not close that boundary. #4377 supplies a counting proof, not a plugged-in evaluator or payment for the executed callback. Outcome remains partial: no first physical repair, new algorithm, new endpoint, new asymptotic tariff, complete CRT runtime/multiplication theorem or κ improvement is claimed.

Evidence

• Existing caller receipt: 29,440 hole-key evaluations, zero good-address evaluations; 133,110,830 repair tape actions and 226,260,150 record-bit-weighted repair actions across ten invocations. Producer 8,422 and scanner 4,548,300 actions separately counted; 2,266,880 key-interface actions are already included in repair. Host key arithmetic and macro/layer adapters are excluded.
• Every reverse key is compared with Audit PR4100 arbitrary-source CRT repair with an explicit inverse #4169 on every address of every layer; every invocation matches the pinned table-based caller oracle. This does not make the inverse formula or fixture endpoint new work.
• Twenty-four detected faults: wrong keys, missing/truncated masks, keying good addresses, skipped sort, every clear and head return, producer geometry faults and corrupt source bytes.
• A stale-source mutant is key-identical on the actual layer and explicitly not counted as a detection; a target-independent parity fixture makes it fail the real hole-key check. Overmarking one good address preserves output but incurs one extra key.
• Fresh consolidation validation: the registered Python 3.12 caller and control commands passed in an author-local environment with outcome=partial in 56.46 seconds, comparing the retained receipts without rewriting them. The prior OpenAI review/optimized replay of the Anthropic implementation and extracted replay remain historical evidence; the new scope/attribution edits also received an independent Anthropic static review. Agent-run evidence, not hosted CI or human/proof-kernel acceptance.
• Reproduce: python3.12 -B scripts/research_package.py --package research/crt-mask-repair-bridge.
• Only Python 3.12 and the standard library are required. Nix/direnv configuration is workstation-local and untracked; .gitignore is unchanged from upstream.

Scope and attribution

The algorithm and CRT tariff remain #2250/#4328's. Jose Manuel's #3055 at 9dc841c9a101435fdd15e0f7623bf16b6d0363e9 already implements encoded current-inverse keys, hole flags, physical stable two-bucket radix repair, destination equality checks and disposal (proof §§4–6, implementation). His #3608 at c6020096322012f7cc41236c3ac15a6b6cd3f8b9 supplies encoded variable-pair sparse XOR control reusing that repair (proof, engine). These are physical-repair prior art in a long-record regime; their resource absorption is not the CRT R = Θ(p) scanner-mask boundary. The residual addition is actual CRT scanner-mask consumption and disposal/head gates, not a general physical-repair novelty claim.

Package research/crt-mask-repair-bridge; historical canonical base f9f85c1e4fa1e955ff39f10fca447e3e2a4e4adc and all source pins/receipts are preserved. At inspected live main fbc9ea0249746acca4f504d4fda2b84e0e36cf0d, the selected result is conditional #4470, κ = 0.00302015, priority revision 7; this package changes no selected frontier. Only unary one-target u2,t2,y8 geometry is covered. The unchanged source manifest retains #4100/#4169/#4201/#4208/#4328/#4377 and PR2250 lineage; #3055/#3608 are exact-head attribution/comparison, not added runtime inputs. Remaining obligations: literal metered key evaluator, macro/inter-layer tape handoff, and many-target/transported geometry.

Prepared for codyw912 with substantial Anthropic implementation and OpenAI coordination/verification assistance, including OpenAI-family scope/attribution edits; no human or formal-kernel acceptance. Credit Cinna Davis (#4100 caller), Hiroki KUMAZAKI (#4169 inverse keys and exact ten-layer fixture), ToikY (#4201/#4208 scanner/producer), Jose Manuel (#2250/#3055/#3608 repair lineage), Ryanson Jonathan (#4328 CRT tariff/handoff), unpaidmoon (#4377 counting proof), and Douglas Colkitt/OpenAI and the retained NOTICE lineage. Apache-2.0.

Current publication revision

Branch research/20261022-crt-repair-bridge, commit 60d0b09f13a9f9ebdd93d98d83d16ce6907d70b6. This follow-up removes agent-added .envrc, devenv.nix, devenv.yaml, devenv.lock and related .gitignore changes from the PR contents, retaining the local files privately ignored. The PR adds only research/crt-mask-repair-bridge/; scientific code, receipts and pins are unchanged. Independent Anthropic review approved the environment-boundary correction. Registered caller/control replay passed again in 56.71 seconds with outcome=partial, and 20-file committed-byte/source-history checks passed at this head. Earlier export/integrity checks belong to commit 1013aa5cfc73. No hosted CI success or human/proof-kernel acceptance is claimed. Package-local LOCAL-ONLY notes retain the pre-publication state; this PR provides the public artifact, not scientific acceptance.

Hosted status

The Registered research packages run at current head 60d0b09f13a9 reports action_required. No hosted success is claimed; upstream approval is required before the fork workflow can run.