[JoyboyBrian]

PR #4795 supplied a complete marker-alphabet key service but explicitly left the binary realization of its opaque-payload repair port open. This package compiles and actually executes that entire retained service using one-bit reads/writes, unit bit-head moves and one-bit Boolean/control operations.

For the fixed u2,t2,y8 geometry and targets 0,1,2, a block-simulation proof covers every legal address and supplied target-independent 32-bit ROM. Each source cell has an injective 13-bit encoding for 4,096 opaque payloads plus eight symbols; tagged registers distinguish Booleans from symbols. Each core source instruction lowers to at most 43 actions. Initializing and erasing 492 control bits adds 984 executed actions, giving 141,346,714 binary actions per key from the retained source bound. Observed counts around 0.9 million are separate evidence, not the proof.

The original bridge.split calls the bit executor at its first two actual marked holes using the same physical counter/bad-stream tapes, then continues to the third hole. There is no call-boundary address decoding or key copy inside the executor. The original one_pass consumes a projection of the completed returned pairs. The facade still decodes symbols for the unlowered outer controller, and closing the captured prefix is explicit verifier setup; this does not claim a full binary repair layer.

Validation:

• Complete 4,104-symbol codec, 8,208 literal read/dirty-write cases and 172 Boolean/alias cases.
• Four standalone full binary key runs plus two actual caller invocations, totaling 5,348,850 bit/control actions. Independent marker semantics checks all 44,340 source boundaries in the transported parity witness.
• Five faults detected: short bit stride, dirty ROM head, omitted control erasure, dropped Boolean tag and marker collision. Three malformed framed packets reject with physical/control cleanup.
• Registered Python 3.12 replay and source pins pass in about three seconds; a separate read-only review corroborates compiler, proof and actual caller scope.
• Publication-byte/source-history, export integrity and Git-free optimized replay results are recorded below.

Scope remains fixed geometry. Full outer producer/scanner/radix/reinsert control, raw macro-S generation, physical ROM supply, encoded variable-width parameters, uniform compiler costs and total E(p,ell) remain open. No full 2,944-call binary layer, historical suite, new κ or selected-frontier change.

Exact predecessor #4795: 4e4cf7e47f3fc1c40e775d84fb2e7cd2603658cd, retained unchanged on canonical base b91956876e5b4122fd89179f1b57557ff8378f15. The source closure also pins #4752's query program/proof, original #4637 bridge/oracle, scanner increment, compact word and notice. Current #4938 covers evaluator/layout caps, #4942 and reservoir work cover interchange wrappers, #4946 is a census audit, and #4965 studies complex stopped recursion; none supplies this binary key executor.

Canonical claim: #4709 (comment)

Prepared by JoyboyBrian with substantial OpenAI Codex assistance, preserving all predecessor attribution and source notices; Apache-2.0. No independent human/kernel acceptance claimed. Hosted CI is pending/unverified.

Publication evidence at head 67154122b2867e6dee5b9b25f2f1c862a3e81e95:

• Three commits separate the literal compiler/proof, independent oracle/caller controls, and reproducible package evidence. Working tree clean; diff whitespace check passes.
• Committed-byte and source-history checks pass for all 17 exported files; canonical base is an ancestor and all nine live inputs are pinned.
• Export integrity passes. Archive SHA-256: 04f1ca0cbf7aa955e2240151c57600c5e3d237d89e8e418ef909e3677d35ea7c.
• Fresh Git-free export replay with Python -I -O -W error -B passes in 2.703 seconds. Independent exported-source drift and frozen-result drift controls both reject.

Final scoped overlap checkpoint: offline search on 2026-10-23 05:16 UTC of binary lowering, binary key, crt-paid-key-service, and opaque payload used the complete 323-open-PR snapshot (05:14:01 UTC), 4,077-discussion snapshot (05:14:50 UTC), and 6,638 local notes at the exact submission head. Thirteen hits were the retained predecessors, this claim/package/hub pointer, and unrelated selected-stream work; no matching new executor lead was found. This is a cached prose search, not a novelty proof. A live recheck confirms merged #4795 still has head 4e4cf7e47f3fc1c40e775d84fb2e7cd2603658cd.

[jwr404]

Post-merge independent check of head 67154122.

• Headline bound. The 3,287,110 factor is Execute paid CRT key service at the actual repair port #4795's proved_instruction_bound in research/crt-paid-key-service/COMPONENT.json. With 2 × 492 = 984 lifecycle actions, 43 × 3,287,110 + 984 = 141,346,714 exactly.
• Replay. It passes on Linux with CPython 3.12.14 in 3.5 s (PASS_LITERAL_BINARY_KEY_PORT, then PASS_REGISTERED_RESEARCH_PACKAGE … outcome=completed). The receipt reports max_core_block = 43 and lifecycle_actions = 984 for all three target programs (106,583 / 106,713 / 106,843 binary instructions).
• Observed versus proved. The executed keys use 859,622 to 914,737 binary actions, about 155× under the proved per-key bound. That fits the README's statement that observed counts are separate evidence and the bound comes from the retained source bound. The source bound itself is loose: one audited key executes 44,340 source instructions against 3,287,110.

I did not re-derive the 43-action block lemma or the codec injectivity proof. This is an independent check, not acceptance, and κ is unchanged.

Reviewed by James Reed with substantial Anthropic Claude (Claude Code) assistance.