Let the time horizon fall from a twentieth to a tenth by move 60

[aywrite]

With no movestogo the share was the clock over 20 at every move, and at 10+0.1 master ended its games with about 3.5 s of the 10 s clock unspent. Now the horizon is 20 at the first move searched on a clock since ucinewgame and falls in a straight line to 10 at move 60. The interface keeps that first move number; an earlier one (a new game without ucinewgame, or a takeback) starts the count again. Everything else in the share is as before.

Results against master 8ade39b:

• 10+0.1: sprt [0, 10] passed, +18 ±10 over 2,000 games (LLR 5.28). A stopped estimate, so it leans high.
• 30+0.3: a fixed 1,000 games, +18 ±13 (95% +5 to +32).
• No time losses in either. The candidate had 1.094 of master's time at 10+0.1 and ended games 1,269 ms lower, against 1,272 predicted by re-running master's clocks. At 30+0.3 it had 1.076 of the time and ended 3,166 ms lower.

The bench is fixed depth and unchanged at 11,088,536. One line in docs/ROADMAP.md named the removed constant and now describes the falling horizon.

🤖 Generated with Claude Code

https://claude.ai/code/session_011saZHGHgVvBWxJ7nougTPN

---

Generated by Claude Code

[github-actions]

Speed against 8ade39b1

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      4218861        4292922   +1.8%
    2      4352191        4240930   -2.6%
    3      4344562        4349006   +0.1%
    4      4305689        4288384   -0.4%
    5      4270646        4246753   -0.6%
    6      4322800        4306936   -0.4%
    7      4323779        4254306   -1.6%
    8      4282577        4322163   +0.9%
    9      4319949        4253029   -1.5%
   10      4360434        4272026   -2.0%
   11      4409412        4265838   -3.3%
   12      4279467        4288384   +0.2%
   13      4364422        4295569   -1.6%
   14      4371515        4378611   +0.2%
   15      4359806        4337128   -0.5%
   16      4286178        4272670   -0.3%
   17      4308543        4323879   +0.4%
   18      4349897        4333914   -0.4%
   19      4305615        4190059   -2.7%
   20      4345643        4238140   -2.5%
   21      4200613        4271364   +1.7%
   22      4224242        4309650   +2.0%
   23      4346684        4326283   -0.5%
   24      4220962        4079101   -3.4%
   25      4250888        4167857   -2.0%
   26      4118506        4355583   +5.8%
   27      4241975        4274137   +0.8%
   28      4418482        4323624   -2.1%
   29      4311699        4392557   +1.9%
   30      4232021        4271972   +0.9%
   31      4174777        4340521   +4.0%
   32      4278861        4241227   -0.9%
   33      4330777        4245343   -2.0%
   34      4300751        4302145   +0.0%
   35      4395916        4325469   -1.6%
   36      4371652        4434137   +1.4%
   37      4413047        4412351   -0.0%
   38      4295745        4447504   +3.5%
   39      4498557        4402125   -2.1%
   40      4441499        4419465   -0.5%

              nodes    time  median nps  faster half
base       11088536  2.57 s     4315824      4372051
candidate  11088536  2.58 s     4298857      4357153
change                            -0.4%        -0.3%

paired change -0.4%, 95% interval -1.0% to +0.2%
diagnostic, on the default layout alone -0.5%, 95% interval -1.3% to +0.6%, 13 rounds

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

Speed: -0.4% (bench nps, 95% interval -1.0% to +0.2%, 40 interleaved rounds over shuffled layouts vs 8ade39b1)

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 8ade39b1, one cachegrind run a side, bench

             instructions     nodes  per node
base       16,519,610,883  11088536    1489.8
candidate  16,519,608,649  11088536    1489.8
change             -0.00%              -0.00%

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