docs(docs): List endgame scaling by material signature as rejected

[aywrite]

Adds one entry to the measured and rejected list in docs/ROADMAP.md.

The scale cut the score for the side ahead in opposite bishop endings and in pawnless endings a bishop or less ahead (Stockfish's fractions for the pawnless signatures). It played +1 ±5 over 8,000 games at 10+0.1 against d82638a, unresolved under sprt [0, 5] (LLR -0.93), and was not merged. The code stays on claude/eloquent-newton-y5ir97 at 2448bad.

Docs only; no engine change.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Be9kcyZC1EuZdKrs41a9Sc

---

Generated by Claude Code

[github-actions]

Speed against 72564490

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      4407357        4278542   -2.9%
    2      4200650        4157957   -1.0%
    3      4162208        4244865   +2.0%
    4      4154629        4156919   +0.1%
    5      4149441        4194340   +1.1%
    6      4285436        4281402   -0.1%
    7      4201953        4104382   -2.3%
    8      4134822        4175268   +1.0%
    9      4283310        4214764   -1.6%
   10      4206249        4263739   +1.4%
   11      4159233        4205073   +1.1%
   12      4199027        4141698   -1.4%
   13      4207401        4166879   -1.0%
   14      4145133        4143072   -0.0%
   15      4319425        4343364   +0.6%
   16      4346952        4190678   -3.6%
   17      4300814        4319693   +0.4%
   18      4244168        4261312   +0.4%
   19      4176177        4149314   -0.6%
   20      4165111        4121450   -1.0%
   21      4295116        4176891   -2.8%
   22      4179864        4221724   +1.0%
   23      4323838        4405574   +1.9%
   24      4280664        4377674   +2.3%
   25      4379301        4392125   +0.3%
   26      4460288        4393561   -1.5%
   27      4455728        4486078   +0.7%
   28      4148157        4187755   +1.0%
   29      4213433        4468877   +6.1%
   30      4279817        4233881   -1.1%
   31      4458229        4332385   -2.8%
   32      4360593        4400841   +0.9%
   33      4336955        4432414   +2.2%
   34      4543545        4473200   -1.5%
   35      4426577        4347540   -1.8%
   36      4077832        4196210   +2.9%
   37      4230018        4269652   +0.9%
   38      4240030        4185612   -1.3%
   39      4247213        4300791   +1.3%
   40      4217520        4228377   +0.3%

              nodes    time  median nps  faster half
base       11088536  2.61 s     4242099      4351766
candidate  11088536  2.62 s     4239373      4353681
change                            -0.1%        +0.0%

paired change -0.0%, 95% interval -0.6% to +0.6%
diagnostic, on the default layout alone +0.2%, 95% interval -1.7% to +1.9%, 13 rounds

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

Speed: -0.0% (bench nps, 95% interval -0.6% to +0.6%, 40 interleaved rounds over shuffled layouts vs 72564490)

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 72564490, one cachegrind run a side, bench

             instructions     nodes  per node
base       14,997,100,316  11088536    1352.5
candidate  14,997,100,310  11088536    1352.5
change             -0.00%              -0.00%

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