[aywrite]

Pass 6 of the architecture review, candidate F1 (pass 5's E2 re-read). A shape change with no search change: the bench is master's 4102609 on both sides, position by position (3868987 before the rebase over the king attack fit), and the tactics and strategy pins do not move.

The decision about a quiet move ordered late was spread through engine.rs: twenty constants, two predicates with six and nine parameters, a staging field on the engine, and two derivations of the features the attention model reads (with a third in fit_attention.py that had to agree). It is now arche-core/src/late_move.rs. The move loop asks one question, decide(&Search, &mut Node, m, searched), and gets Skip or the plies to scout. Search is three references (board, ordering, config), so the module names what it reads and a test builds one with no engine.

The gate and the reduction ledger now read the features from one function over the same node. The staged field goes: the staging travels as an Option<&Staged> parameter through search_child and windowed, so a reduced move inside a scout cannot consume its parent's staging. The ledger's history_max column becomes the node's held denominator, as the gate always read it, rather than a fresh walk per staging; INSTRUMENTS.md says so and a test holds the cache.

Callgrind on the identical bench: +0.077% instructions, the per-move Node literal against a cheaper parameter, under the clock floor. Tests 475 to 478 debug, 473 to 476 release.

The handoff with the eight grilled decisions is in the planning repo (handoffs/late-move-module.md). One Opus review round, both sides rebuilt: no behavioural defect, nine findings folded into the commit.

🤖 Generated with Claude Code

[aywrite]

Review round 1 (Opus reviewer, own build directory, both sides rebuilt from git archive exports with tar -m; bench 3868987 at 4b1b80f and at the candidate, position by position; the decision path read line by line against the old reduces and model_gate; mutations on the clamps, the threshold order and the history/denominator swap all caught by tests). No behavioural defect. Nine findings, all folded into the one commit:

1. The armed instrument path paid an eval at depth-3 nodes the base never paid, because features reached the eval through the same lazy pair as the denominator. The pair is now two lazies; features asks for the denominator alone.
2. Two comments claimed the features were "read back off the node" and "derived once and carried"; they are derived by a second call of the same function. Both now say the real property: one derivation site, so the two readings cannot diverge.
3. The ledger's history_max column changed meaning (the node's held denominator, as the gate reads it, rather than a fresh walk per staging) and neither the commit body nor INSTRUMENTS.md said so. Both do now.
4. Nothing tested that the denominator is held for the node. A test now teaches the history between two readings and asserts the second is unchanged; it fails when the cache is bypassed.
5. The new test's closing assertion only checked determinism. Removed; the solved-edge assertions above it are what hold the gate to the recorded row.
6. scripts/tests/test_fit_attention.py still named engine.rs for the clamp. Fixed.
7. The speed paragraph never said which side was faster. Rewritten with the direction and the callgrind attribution (+3,110,295 Ir, alpha_beta +1.9M for the per-move Node literal, windowed -1.6M for the parameter).
8. A 79-column line in ARCHITECTURE.md after the rename. Rewrapped.
9. The implementer's deviation note undercounted the pub(crate) constants (five, not one). Note only; the code was right.

Every number in the commit body was checked against the diff by the reviewer and holds.

[github-actions]

Speed against 7991f406

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     10733042       10678454
    2     10818945       10717789
    3     10721150       10640870
    4     10721290       10527149
    5     10753720       10637890
    6     10624584       10612574
    7     10689862       10524070
    8     10710849       10600454
    9     10578996       10539481

             nodes    time  median nps  fastest nps
base       4102609  0.38 s    10721150     10818945
candidate  4102609  0.39 s    10612574     10717789
change                           -1.0%        -0.9%

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: -1.0% (bench nps, 9 interleaved rounds vs 7991f406, spread 2.2%)