[aywrite]

The occasional whole-tree sweep, run against 7e6ddd7. This revision of it reviews structure and contracts before prose, so the findings are ranked by what a reader or a conclusion loses rather than by how many stylistic hits there were. Seven commits, each reviewable on its own.

Two correctness defects, each its own commit.

• build_corpus.py --sealed was read, validated and then dropped: corpus() took the set and never passed it to Entry, so rows were labelled by the key split while the fit held out the named group. That is the exact mismatch both halves of the tuner document as the leak the shared module exists to prevent, and following the documented procedure produced it with exit 0. One argument, plus a test that reads 1.0 instead of 0.5 without it.
• The published image's version label was the string version.workspace = true. grep -m1 '^version' Cargo.toml finds that line in [package], five above the real version under [workspace.package], and it holds no quote for cut to split on. Two other workflows read the same fact with tomllib; this now does too.

Documentation drift. The readme still described the evaluation as material plus piece square tables; it is material, tables, mobility, king safety and pawn structure. Two documents still called mobility a rook count, which it stopped being in the commit below this branch. The roadmap said whether a fit on our own games buys strength was unsettled, where four fits have each passed an sprt since. The architecture map counted four tuner scripts as three, described a split by game that has been by pair for some time, and never listed src/command.rs. The roadmap now also carries the Miri exposure that arche-core/Cargo.toml says it carries.

Comments a reader could act on and be wrong. ordering.rs said the killers and the history score are read by the cutoff census; they are read by model_gate in the playing search, which is where late move pruning and the deep reduction get their score. Three comments still described the mobility skip as active after the refit priced all four kinds. One comment said "both" of three terms.

One new check and two small fixes. The tactical suite now checks that a bm names a move the generator offers, which the strategic suite has always done; all 300 pass, and sabotaging one fails it. The last unpinned action in twelve workflows is pinned, in the job that gates master. A comment paragraph is moved back over the engine it describes.

Behaviour is preserved throughout except in the two defects above, which are the point. The bench reads 3,868,987 on every commit, the count the parent states. Both test runs pass including the debug one, clippy under -D warnings and fmt are clean, and the 264 script tests pass alongside the one added.

🤖 Generated with Claude Code

[github-actions]

Speed against 7e6ddd71

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      8043765        7864593
    2      8220728        7927521
    3      8123857        8045003
    4      8291586        8478206
    5      8608365        8254138
    6      8199959        8554179
    7      8276067        8304507
    8      8465184        8498934
    9      8276687        8424320

             nodes    time  median nps  fastest nps
base       3868987  0.47 s     8276067      8608365
candidate  3868987  0.47 s     8304507      8554179
change                           +0.3%        -0.6%

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: +0.3% (bench nps, 9 interleaved rounds vs 7e6ddd71, spread 8.3%)