[aywrite]

Four small items, none of which plays a game. The bench is unchanged at
3868987 and the tactical and strategic totals are unchanged, so nothing
here is a play change.

The transposition module said fail low nodes are not stored. They
are. A node where no move raised alpha takes the else branch in
alpha_beta and record_ceiling writes it as an upper bound with the
move that came closest on it. The rest of that sentence is true and is
kept: a cutoff no move can be attributed to, which is what null move
pruning or a static cutoff would want to record, still has nowhere to go
while an entry must name a play.

A floor under the tactical suite, held apart from the exact count.
The exact count is a tripwire. A change that means to move it updates the
number and the tripwire is rearmed one notch lower, and nothing in that
stops the count walking down a few positions at a time until the suite
says nothing. The floor is 210, kept strictly under the snapshot by a
build assertion so that lowering the snapshot past it fails the build
rather than waiting for the suite job. The count has been 243, 241, 240,
237, 236, 229, 221, 226, 227, 228, 231 and 224 since the suite was first
gated and the largest single step in that list is eight, so the floor is
not reached by an arm spending what an arm has ever spent.

Beside it, a list of losses taken on purpose, each naming the position,
what bought it and the version its acceptance runs out at. It ships
empty: the suite misses seventy six positions and none has been read one
at a time, so anything in it now would be a guess and not a record. The
next change that lowers the count is the first to write in it. What makes
writing in it mean something is the checking: an id not in the suite, an
id twice, an entry with no reason, an entry whose version has passed, and
an entry the suite has started finding again. A made up list carrying
each fault once is what says those checks can fail.

Transposition tests that leave ply zero. All twenty five ran at ply
zero, where both halves of the mate score conversion are the identity, so
not one of them could tell a working conversion from a broken one. The
new test stores a mate nine plies down and reads it back two plies down,
for a mate each way and for an ordinary score that no ply moves.

The instrument rate test could not fail on half its name. All three
DEFAULT_EVERY constants are a thousand, so swapping two of them at the
call site left the test passing and the whole workspace with it. Which of
three equal constants a caller named is not checkable by value while they
are equal, and the doc comment now says so. What is checkable is that
sampling returns the rate it was handed, so the three commands are
handed three rates that differ.

Each of the three test changes was checked by injecting the fault it is
meant to catch and watching it fail: the mate conversion flipped either
way, sampling made to ignore its argument, and the suite run dropped
under the floor, where the floor's message fires ahead of the snapshot's.

🤖 Generated with Claude Code

[github-actions]

Speed against e282f255

Both sides built and run on this runner in this job, the way
scripts/speed.sh measures a perf commit. The change is between
medians, and the spread beside it is what the change has to be
read against. 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
    1      9793341       10082418
    2      9492769        9982782
    3      9977067        9635347
    4      9848985       10157514
    5      9596296        9529618
    6     10109474        9738224
    7      9813163       10090965
    8     10343751        9743522
    9      9878937        9989406

             nodes    time  median nps  fastest nps
base       3868987  0.39 s     9848985     10343751
candidate  3868987  0.39 s     9982782     10157514
change                           +1.4%        -1.8%

the change is inside the spread, so the medians make no claim; the
fastest column is the steadier comparison when the machine was not
quiet

Speed: +1.4% (bench nps, 9 interleaved rounds vs e282f255, spread 8.6%)