[aywrite]

Adds three entries to the end of "Measured and rejected" in docs/ROADMAP.md. They cover four arms measured since the list was last updated:

• Multi-cut at a singular test node. +5 ±7 over 4,000 games at 10+0.1 against 048dfa7, unresolved at the cap. Its fixed-node screen read +2 ±21.
• Two refits of the late move pruning skip as a logistic model over the move. With a free depth weight it read −15 ±15 over 1,000 games. With depth tied to the margin's line it read −9 ±12 over 1,500. Both are sprt [0, 10] fails against the margin in 93e5b1b.
• Ordering the zero-scored quiets by their evaluation change. −4 ±9 over 3,000 games, and −6 ±9 over 2,500 with the piece square change alone. Both are sprt [0, 10] fails against 2d7fe54.

Docs only. No code, bench or test changes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01A57SbXV3vACDuNUX5zSKPQ

---

Generated by Claude Code

[github-actions]

Speed against 5d1bb6ef

Measured on AMD EPYC 7763 64-Core Processor.
Both sides built with rustc 1.99.0 (b940084d7 2026-09-28).

Both sides built and run on this runner in this job, the way
scripts/speed.sh measures a perf commit. Each round runs
both sides on a layout of its own, the same compiled code with
its code and data shuffled and moved, so the 95% interval carries
where the code landed as well as the run. The verdict holds the
interval against a 1% threshold. The default layout, the one a
release ships, is measured after as a diagnostic. The node
counts are the search's: they move when the search does, and a
speed change leaves them alone.

round     base nps  candidate nps  change
    1      3964076        4048197   +2.1%
    2      4013856        4090373   +1.9%
    3      3976323        3985575   +0.2%
    4      3957763        4031419   +1.9%
    5      3942491        3888025   -1.4%
    6      3917179        3865183   -1.3%
    7      3962951        3933721   -0.7%
    8      3981450        3968999   -0.3%
    9      3954865        3994846   +1.0%
   10      3974921        3961773   -0.3%
   11      3960737        3918477   -1.1%
   12      3907845        3989721   +2.1%
   13      3887991        3953634   +1.7%
   14      4008398        3953270   -1.4%
   15      4087249        4034733   -1.3%
   16      4070277        4055899   -0.4%
   17      4038268        4036852   -0.0%
   18      4035213        4029002   -0.2%
   19      4023939        3956009   -1.7%
   20      3991392        3996967   +0.1%
   21      4044215        3989380   -1.4%
   22      4080970        4047258   -0.8%
   23      4065514        4046243   -0.5%
   24      4121047        4105996   -0.4%
   25      4051243        4109339   +1.4%
   26      4082310        4063897   -0.5%
   27      4134953        4043193   -2.2%
   28      4024007        4035091   +0.3%
   29      4033359        4018719   -0.4%
   30      3975626        4030648   +1.4%
   31      3968170        3910200   -1.5%
   32      4050486        4068520   +0.4%
   33      4070835        4028361   -1.0%
   34      4010168        3984906   -0.6%
   35      3956153        3987382   +0.8%
   36      4040294        4032262   -0.2%
   37      3987226        3959338   -0.7%
   38      3929726        3998452   +1.7%
   39      3966581        3964507   -0.1%
   40      3876046        3921752   +1.2%

              nodes    time  median nps  faster half
base       11088536  2.77 s     3999895      4054330
candidate  11088536  2.77 s     3997710      4047723
change                            -0.1%        -0.2%

paired change -0.1%, 95% interval -0.5% to +0.3%
diagnostic, on the default layout alone +0.1%, 95% interval -0.4% to +0.6%, 13 rounds

no change beyond ±1.0%: the whole interval is inside it

Speed: -0.1% (bench nps, 95% interval -0.5% to +0.3%, 40 interleaved rounds over shuffled layouts vs 5d1bb6ef)

The instructions each side's bench executed, counted under
cachegrind. The count repeats to within a few hundred
instructions, so a small change here is a real one, but it
prices instructions only: cache misses, mispredicted branches
and where the code lands are the speed's to show.

instructions against 5d1bb6ef, one cachegrind run a side, bench

             instructions     nodes  per node
base       16,519,610,883  11088536    1489.8
candidate  16,519,610,926  11088536    1489.8
change             +0.00%              +0.00%

within ±0.7%, as far as edits not made for speed have moved the count