[jmbarrancoml]

PR2160 supplies the selected rank-42 normalizer's algebra and leaves its prefix-controlled ADDs as a physical implementation obligation. I execute those ADDs as sequential payload permutations, compute their coefficients from full parent matrix headers, and give a common adapter schedule with explicit preparation costs.

The selected object remains idle child 6, copy 213, port 102, with projector diag(0_22,I_42). For five matrices agreeing modulo 5, one selected shear coefficient takes values 1, 21, 16, 11 and 6 modulo 25. The physical replay moves opaque records to those target positions while preserving parent headers, unused stock and middle/suffix labels. A low-only evaluator moves them incorrectly.

I retain 15,332f ordered adapter slots, including zeros, with zero ordinary interchange calls by the adapters. The caller still supplies one weighted child of width 42f. I also implement a two-level good-ancestor evaluator with coordinate-projector labels; it recomputes the full ancestor pivots and checks the coarse-width/index rule. Variable diagonal units stay in the weighted child.

The written bill keeps control preparation explicit:

T_adapters <= N*V*(C + 2*E_max/q^w),  N=15332f.

I derive the linear O(V*n) toll for the stated stopped-recursion domain using the root width and bounded ancestry. I do not assert that bound uniformly for arbitrary independent current width, atom width and descriptor depth. More general ancestor access keeps its actual cost in E_max.

Local verification passed:

python3.12 scripts/research_package.py --package research/selected-bit-controlled-streams
python3 -I -O -B research/selected-bit-controlled-streams/verify.py
git diff --cached --check

The replay checks 51 full address fixtures, all 128 columns of one dense weighted operator, and 1,277,550 positive-case payload bits. Controls reject low-only current and ancestor evaluation, a wrong complete normalizer endpoint, and an inconsistent ancestor column width. Width-one shifts, zero/nonunit coefficients, variable ancestor weights and scratch cleanup receive explicit checks. The payload and ring-operation counters are separate; I claim no measured full Turing-machine execution or hosted CI pass.

This adds only research/selected-bit-controlled-streams/, starting from canonical main 0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2, priority revision 4. The selected frontier and frozen inputs stay unchanged. Discussion2165 contains the scoped claim and handoff. SOURCES.json pins inputs and predecessor heads; the package imports no predecessor executable.

The remaining work includes the weighted callee, further local-frame and mixed fallback descriptors, complete correlated-helper hosting, fresh-field cylinders, COPY/ERASE lifetimes and outer transfer. No new supplier or exponent claim.

Prepared by Jose Manuel (@jmbarrancoml) with substantial OpenAI Codex assistance. I retain commdata2338's PR2160 normal-form/source-binding credit, the retained ordered-affine and source527 methods, Ross Richardson's PR2159 fallback work, Nick Petalas's carry/stock work, Adam Gall's fresh-field/stock analysis, and the selected BIT lineage of eumemic, DreamingOfClouds, DaysSky, Zhihao Chen, Dylan Crow and Jacob Sussman. Apache-2.0; no independent human review claimed.