[aywrite]

Two documentation commits, no change to behaviour (bench 11,088,536 unchanged).

• docs/ROADMAP.md: three arms measured since the rejected list was last written, each with its match figures. They are the depth veto with its time spent (−26 ±20), history in the late move reduction (−2 ±10) and a time horizon from a moves-left estimate (−6 ±12 at equal time). Multi-cut's entry gains its continuation to 5,000 games (+2 ±7, failed).
• src/time_control.rs: the comment on the falling horizon now says that 20 and 10 are a planning horizon, about 0.3 of the moves a game has left, and why. The search spends about 70% of each share, and matches favoured spending faster than an even spread.

🤖 Generated with Claude Code

https://claude.ai/code/session_011saZHGHgVvBWxJ7nougTPN

---

Generated by Claude Code

[github-actions]

Speed against 54fb9a68

Measured on AMD EPYC 9V74 80-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      4125641        4042977   -2.0%
    2      4107872        4084202   -0.6%
    3      4062815        4085400   +0.6%
    4      4054364        4039457   -0.4%
    5      4050458        4262541   +5.2%
    6      4120271        4231469   +2.7%
    7      4006609        4031546   +0.6%
    8      4133766        4166033   +0.8%
    9      4176141        4096044   -1.9%
   10      4287678        4135286   -3.6%
   11      4090741        4115384   +0.6%
   12      4044457        4098517   +1.3%
   13      4073224        4027701   -1.1%
   14      4071004        4070832   -0.0%
   15      4253402        4161385   -2.2%
   16      4243207        4255350   +0.3%
   17      4248495        4229550   -0.4%
   18      4169806        4177089   +0.2%
   19      4117483        3917990   -4.8%
   20      4188723        4179889   -0.2%
   21      4063251        4071307   +0.2%
   22      3918721        3894190   -0.6%
   23      4060898        4071802   +0.3%
   24      3996571        3982374   -0.4%
   25      4130716        3913402   -5.3%
   26      4092396        3942206   -3.7%
   27      4202885        4228885   +0.6%
   28      4028096        4209463   +4.5%
   29      4092340        4085756   -0.2%
   30      4186711        3998567   -4.5%
   31      4116451        4146770   +0.7%
   32      4003494        4117552   +2.8%
   33      4093742        4096703   +0.1%
   34      4295173        4044578   -5.8%
   35      4149405        4007914   -3.4%
   36      4166188        4105352   -1.5%
   37      4095585        4012488   -2.0%
   38      4095199        4031036   -1.6%
   39      4029557        4035925   +0.2%
   40      4034913        4032460   -0.1%

              nodes    time  median nps  faster half
base       11088536  2.71 s     4095392      4175780
candidate  11088536  2.71 s     4084801      4159221
change                            -0.3%        -0.4%

paired change -0.5%, 95% interval -1.3% to +0.1%
diagnostic, on the default layout alone +0.4%, 95% interval -0.4% to +1.8%, 13 rounds

not resolved: the interval reaches past ±1.0% without clearing it, so
more rounds are needed to say either way

Speed: -0.5% (bench nps, 95% interval -1.3% to +0.1%, 40 interleaved rounds over shuffled layouts vs 54fb9a68)

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 54fb9a68, one cachegrind run a side, bench

             instructions     nodes  per node
base       16,519,609,125  11088536    1489.8
candidate  16,519,609,168  11088536    1489.8
change             +0.00%              +0.00%

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