Every zip() now says what it expects, and two of them did not

[qmohsu]

zip() without strict= truncates to the shorter argument and says nothing.
BASELINE in tests/test_lint_debt_only_shrinks.py recorded 41 of them, and
the obvious move is ruff --fix --unsafe-fixes. It writes strict=False —
the current behaviour, spelled out loud. That takes the count to zero, looks
exactly like the job being done, and preserves every defect the rule exists to
expose. Done by hand instead, one site at a time.

58 sites, not 41

The other 17 are in notebooks/, which is not in SOURCE_DIRS, so no ruff
guard here had ever read them — in the three notebooks a student is most likely
to open. Same shape as KNOWN_NON_EXAMPLE_CHAPTER_FILES: a file outside the
sweep's scope is a file nothing sweeps, and that scope is a tuple somebody wrote
once.

One real defect

ch5_fingerprinting/example_comparison.py built colors_box with five shades
against six methods, so Linear Regression's box kept matplotlib's default blue
among five pastels and read as deliberately highlighted. Nothing raised, nothing
left a range, no printed number moved — the sixth box simply never got its
colour, which only opening the PNG shows. comparison_all_methods.{svg,pdf,png}
regenerated, and all three reproduce byte-identically from the fixed code.

One site where strict=True is wrong

zip(scales, scales[1:]) is the consecutive-pairs idiom and depends on the
truncation, so strict raises on every call — confirmed by running it. It is
itertools.pairwise now. zip(xs, count()) and zip(xs, cycle(ys)) are the
same family; the tell is a second argument derived from the first, and a grep
for [1:] finds only the spelling you thought of.

The ratchet is tighter, not looser

B905 is deleted from BASELINE rather than set to 0, and the deletion
is the guard: test_no_rule_has_more_findings_than_it_did fails on any rule not
listed. Verified by adding one unguarded zip() and watching
B905: 1 findings, new, not by reading the assertion.

A green sweep only means something where the line runs

Measured with coverage: 41 of the 42 .py sites execute. 34 run under
the examples, the generators, and two paths a plain run never reaches — ch4's
--data branch and ch7's --animate callback, which between them hold two
sites nothing else touches. The other seven — six inside test bodies, one in
core/fingerprinting — needed the harness fixed before they could be seen:
[tool.coverage.run] omit in pyproject.toml excludes */tests/*, so a run
measuring test files reports every line in them as never executed, which looks
exactly like "this code never runs". With a scratch rcfile that drops the omit,
all seven execute.

The one remaining site is ch8's unobservable-mode print, behind
if n_unobservable > 0 — false for the fixes-aided configuration by design,
which is the point the example makes. Its twin in the odometry-only branch does
execute, and the pairing is equal-length by construction either way. The 17
notebook sites are covered by tests/docs/test_notebooks_run.py.

Test plan

• ruff check --statistics over SOURCE_DIRS: B905 gone, every other rule
  unchanged against BASELINE
• ruff check --select B905 .: zero repo-wide, notebooks included
• Full suite: 3404 passed, 21 skipped (24m37s) — same counts as main
• tests/docs/test_notebooks_run.py: 14 passed, all seven notebooks execute
  in a real kernel
• All 14 example invocations and both generators exit 0 under strict=True
• black --check: 299 files clean (six files reflowed by the inserted
  keyword, and every changed line in them mentions zip/strict)
• Mutation: one unguarded zip() added → ratchet red; removed → green
• ch5 figure regenerated: svg/pdf/png byte-identical to what is committed

🤖 Generated with Claude Code