[JoyboyBrian]

The actual raw PR5402 channel comprising every full-frame history value and all eight partial-frame endpoint caches cannot determine the coupled-query consumer's b=ZQY_return=ZRy on arbitrary independent inputs. A source-bound 104-role displacement leaves that entire channel unchanged while changing target mode one by -2 K_m,1^-1 R v for any field v. Coefficientwise cancellation in the free-word algebra proves this uniformly at every positive tensor width; finite-width rank is not the proof.

This resolves the endpoint-cache case excluded by PR5447. Each cache defeats an earlier fixed-helper kernel direction; the new witness uses 57 X, three Y and 44 dirty helper inputs. The proof distinguishes original helper retractions, which could conditionally enable decoding, from post-word retractions that vanish identically. It binds the actual consumer via all 112 phase-aware input wrappers and explicitly retained provider identities.

Validation: complete 641-input symbolic replay and an independently implemented direct-displacement replay pass normally and under -O; six on-disk corruption controls pass. Committed source bytes/history, official export validation, and all five registered commands in a read-only Git-free export pass with unchanged files. A separate Codex agent independently checked 57,344 wrapper/address actions and the proof; this is not human or proof-kernel acceptance.

Scope: no complete-machine blindness, physical observer implementation, paid-return improvement, price lower bound, finite-precision theorem, new kappa or selected-frontier change. Other partial histories, lower-frame main observations and correctly versioned original helper information remain open interfaces.

Canonical investigation: #5488. Preserves PR5402/provider, PR5427/5447 history, cache/retraction and PR5415/5419 consumer attribution in the package. PR5489's separate local helper transport contract is distinct. JoyboyBrian with substantial OpenAI Codex assistance; Apache-2.0.