Record that one P104 encoder fiber equals axis-0 U.

[unpaidmoon]

Change and mathematical scope

Partial audit, no new κ. Selected κ stays 0.001672660197076918.

On the encoder emitted by the frozen PR #3979 builder bytes that PR #3996 vendors, the bank-0 column-0 fiber is axis-0 multiplication by U = [[I_8, B/2], [0, I_96]]. The 576-event program occurs once, at offset 41600, inside the 203008-event encoder. Interpreting those events gives U, and not U^{-1}, U^t, or U^{-t}. B is rebuilt from the canonical P104 fixture and matches the published numerator digest.

This is one column of the row factor U_r inside L = U_r U_c^{-t}. It is not a correction program, a physical consumer, or a supplier admission.

Task and overlap

• Base: bb4bccbf42844d2801487450727c2f63e10d1d0f (canonical main, priority revision 6).
• Head: e8861a715b3d22fbf398393a98604caaa73c4f8d.
• Task: p104-side-correction, scope changed.
• Handoff: discussion Handoffs and requests for independent checks #1166 comment 19035716. PR Audit the P104 index-to-tensor bridge without replaying events. #4125 checked the index bridge and did not read the events. This package is that one-fiber comparison.
• Builder SHA256 27bae82fe247fe49b8fc3b5efbd429e2c8c0ee8f56cc4ace748cbb681a5abacc. PR Audit every scalar event in the actual P104 tensor code #3996’s checker is not imported.
• Fork branch: cursor/p104-axis0-encoder-fiber-f55c.

The eligible list has the 12 entries above. Item 8 is the bridge; this PR is the unread fiber, not a second bridge audit.

Validation

python3.12 -B scripts/research_package.py --package research/p104-axis0-encoder-fiber

PASS_P104_AXIS0_ENCODER_FIBER with match_offset 41600 and map U. PASS_REGISTERED_RESEARCH_PACKAGE ... outcome=partial. Source-history check passed. The same commands pass under python3.12 -O.

Failure controls: a flipped add sign misses U; column 1 is not this slice; dropping the restoring doubles leaves a tail scale of 1/2.

Not run: the other encoder fibers, PR #3996’s checker, the 763,392-event census, Lean, Docker, or the selected cold replay.

Credit

Chris R / Clear1nsight, PR #3979, for the encoder and builder. ToikY, PR #3996, for vendoring those bytes and for the fiber question. PR #4125 for the index bridge this package does not repeat. unpaidmoon, with substantial Cursor Grok 4.7 assistance. Apache-2.0.

Open next questions

1. Do the other 103 bank-0 axis-0 fibers differ from column 0 only by the column index?
2. Does one repeated consumer of (N-P) ⊗ M pay less than rebuilding the uncoded Gram pair? Blocked: no consumer or complete allowance is named.