perf(ast): exclude_tests is quadratic in depth x one-attribute-item count

[dekobon]

Summary

--exclude-tests over Rust is quadratic in depth × attributed-item count when
each attributed item carries a run of one attribute. PR #1446 fixed the
sibling shape (one long run, many attributes) by reusing a verdict across the
run; runs of length 1 have nothing to reuse, so this axis is untouched.

Roughly half of the cost predates #1446's attribute arm: it is
rust_outer_attr_marks_test restarting at the parent's first child once per
item, i.e. the #1100 budget design rather than any recent regression.

Reproduction

D nested fns, then 3*D separate one-attribute items at the innermost
level:

lines  = ["fn f%d() {" % i for i in range(D)]
lines += ["#[cfg(test)]\nfn t%d() {}" % i for i in range(3 * D)]
lines += ["}"] * D

$ bca metrics --no-config --jobs 1 --metrics nom --exclude-tests -p dN.rs

D	--exclude-tests	flag off	with #1446's attribute arm disabled
250	0.20 s	0.00 s	0.09 s
500	0.75 s	0.01 s	0.37 s
1000	3.07 s	0.03 s	1.38 s

~4× per doubling in all three columns that grow, so the class is quadratic with
and without the attribute arm.

Root cause

forward_attribute_scan_budget is max(6, 3 * depth)
(big-code-analysis-ast/src/checker.rs). At depth D it permits a width-3D
parent onto the forward reading, and rust_attribute_run_under then walks the
parent's children from index 0 for every item it is asked about. With 3D
items under one parent that is O(D²).

The budget scales with depth because a previous_sibling climb costs O(depth)
while a cursor step is flat (#1100's measurements, recorded in the constants'
doc comments). That reasoning is sound per call; what it does not price is
being called O(width) times on the same parent.

Fix sketch

The shape of #1446's fix does not transfer directly — its reach works because
every attribute in a run shares one answer, and here each item has its own.
Options:

1. Reach on the other axis. A verdict for item i could report that it
   answers for the run boundary it already walked past, letting the next item
   start where the last stopped. Needs the walker to carry a per-parent cursor
   rather than a single reach, which is more state than fix(metrics/loc): six LOC fixes from the 2026-09-10 residue #1446 added.
2. Cap the forward budget by absolute width, not just depth. A parent wider
   than some constant takes the backward reading regardless of depth. Cheap, but
   it re-opens the O(depth²) case perf(checker): the exclude_tests attribute scan is O(attributes x depth) #1100 closed for deep-and-narrow trees, so it
   needs that fixture re-measured.
3. Memoize per parent. One (parent_id, run_verdict) slot on the walk would
   collapse the repeat asks. Adds walk state and a cache-invalidation question.

Measure before choosing: the shape is pathological, and real Rust is a few
percent attributed. This is filed because the class is real and now named, not
because it has been seen in practice.

Found while fixing the #1446 review's HIGH finding, by checking whether the fix
covered the adjacent diagonal. It does not, and the new nom/deep-attribute-run
probe does not cover it either.