Scoped negative result: SW4 pre-read insertion slides do not change the complex cost (seed 20)

[navinor]

Scoped negative result. No new kappa, no certificate change. Only research/complex-insert-slides-seed20/ is added. The selected record, upstream/ and every supplier file are unchanged.

Base: f567884aeb8851b6e794af1184119a3bfbe1e631. Complex-supplier track (20-39), OS-random seed 20.

Claim

SW4 (PR 699, complex word of the selected frontier ancestry) inserts each r2h compensation pre-read add(t,h,-c) at one fixed position ("right after h's last existing pre-read"). Any such pre-read may instead sit anywhere before its helper's copy gate, where h = E_h, without changing the exact B-end state. No placement beats the pinned one.

Complete deterministic census on the pinned promoted P6k4 word (content 0390ec65..., refused by build_sw4.py if different), with the 16 recipe moves otherwise unchanged:

• each window is [0, copy_index(h)] in the deletion-applied B list (637-830 positions), identified up to positional equivalence (positions between consecutive gates touching the moved target or helper): 7 classes per pre-read;
• all 56 single slides, all 1,372 class-pair slides, and both orders when one helper's two pre-reads land together (4): 1,432 configurations;
• every configuration preserves the exact B-end state mod p (all 728 registers); 1,060 pass the inherited frame refit/profile checks; every legal configuration has exactly the pinned paid profile {1:2548, 2:342, 3:438, 4:186, 5:56, 8:8} (rank mass 5,634, pinned widths), so the complex endpoint is pinned at 0.0016634287545401745; the 372 refused configurations are all frame-cap violations. 0 improving.

Insertion position is invisible to cost whenever it is legal: the pinned choice is one representative of a plateau, not a tuned parameter.

Pricing and controls

• Priced by the exact complex endpoint functional of the pinned moment model: root of sum_d n_d*d*(73/d)^a = 73*7168 over H = 72*H_local + {1:1, 3:1, 44:1, 48:2396}. Calibration: on the frame8 profile it reproduces the selected record's complex endpoint 1663966829367501881629/10^24 to float precision; the pinned exact mercator/exponential engine (engines/moment.py, unchanged) brackets every reported root within 2e-15 with exact rational residuals.
• Controls: the pinned schedule through this package's placement code reproduces the pinned SW4 output word byte for byte (d232183f...) and profile; an out-of-window slide (past the copy gate) is rejected by exactness; six behavior/failure tests pin byte-exact reproduction, the class partition, out-of-window inexactness, the absence of scale-of-target hazards in every window, in-window exactness, and the endpoint functional.

Reproduce

python3 -B research/complex-insert-slides-seed20/verify.py

Expected final line (about 8 minutes, Python 3.11+, stdlib only, no optimizer/network):

PASS_INSERT_SLIDES_SEED20 singles=56 pairs=1372 swaps=4 legal=1060 improving=0 best_root=0.001663428754540

verify.py re-runs the tests, re-enumerates the complete census, re-runs every control and compares counts, roots, the exact-engine cross-check and the canonical rows digest against EXPECTED.json. Not run: make verify, make frontier799-verify, Linux CI, Lean (no generator, certificate or selected-record file is touched; scripts/verify_selected_result.py passes).

Scope

Finite scoped negative result, not an impossibility theorem. Not covered: triples and larger slide groups; relocating the deleted h2r entrance pre-reads (deletion is exactness-forced); the converted reads themselves; relaxed frame floors, moved helper boundaries, other decoders and other words; frame assignment of the fixed program (closed by PR 815 and #864); read-set choice (PR 775, #876, SW4's own census); owners, hosting, the invoice, the bit supplier and the outer theorems.

Prior and concurrent work

• BRoute / Jeremy Kothe's SW4 (PR 699) and its search record vary read subsets, not insertion positions; construction credit and lineage are in NOTICE.md/SOURCES.json.
• PR 775 (coupled partial read exchanges, seed 21) excludes "alternative insertion positions" in its scope list; this package completes exactly that exclusion with a different method (positional-class enumeration).
• Scoped negative result: simultaneous full pair/triple exchanges do not beat SW4 (seed 29) #876 (seed 29) covers simultaneous full pair/triple exchanges; its vendored P6K4 input is byte-identical to the pin used here (acknowledged).
• PR 815 samples commuting reorders by annealing; this census is complete over a precisely defined reorder subfamily and agrees with its picture.
• Scoped negative result: no frame reassignment of the selected X4 complex word admits the next tick (exact duals, no new kappa) #864 (seed 28) excludes frame reassignment of the selected X4 word by exact duals; gate positions were not in its scope.

Prepared for @navinor with help from opencode. Apache-2.0. No independent human review, no kappa change, no formal proof claimed. Complex-supplier track, OS-random seed 20; claim recorded in research/complex-insert-slides-seed20/CLAIM.md (no pinned sprint coordination issue exists).