Raise the cognitive-complexity threshold to 30 and clear the rest

[helly25]

Third clang-tidy triage PR. Clears all 37 readability-function-cognitive-complexity findings, touching 6 files instead of 20.

The threshold was not the real problem

Of the 37 findings at threshold 25, only 9 were production functions. The other 28 were gtest TestBody, where the complexity is ASSERT_* macros expanding to branches — not logic anyone should refactor.

threshold	production over	tests over
25 (was)	9	28
30 (now)	5	22
35	2	20
60	0	3

No threshold fixes the tests — TestBody reaches 122 — and one high enough to clear them makes the check decorative for production code. IgnoreMacros looked like the obvious lever but measured too blunt: it silences all 37 including the genuine ones, because it excuses any function touching a macro rather than discounting macro-contributed complexity.

Threshold 30, and tests handled once

Threshold 30 clears the four borderline production functions (26, 29, 29, 30) that are dense-but-fine, while still flagging 32/34/35/39/54.

Tests are relaxed in tools/clang_tidy.sh, which lints *_test.cc with the check appended as disabled.

This cannot be expressed in .clang-tidy: its schema is exactly Checks, CheckOptions, WarningsAsErrors, HeaderFilterRegex, ExcludeHeaderFilterRegex, HeaderFileExtensions, ImplementationFileExtensions, FormatStyle, SystemHeaders, User — there is no per-file-pattern section, and its only granularity is per directory, which cannot separate tests that live beside the code they test.

Doing it in the hook means one statement of the rule instead of the same NOLINT comment on 24 tests, and new tests are covered without anyone remembering to annotate them. --checks appends to the configured Checks (verified), so tests keep every other check.

The five production functions, case by case

• mope_main.cc Process (39) — split. Its complexity was almost entirely one nested block parsing --set=<key>=<value> (a section-path loop inside a branch inside a loop). Extracted as ApplySetFlag, leaving Process a linear pipeline.
• FindMiddleSnake (54) — suppressed. The forward/reverse D-path search is one algorithm from Myers' paper; the two diagonal windows, parity rule and overlap test only make sense read together.
• AppendSideBySide (35), AppendContext (34) — suppressed. The branching is the diff output format — row kinds, gutter markers, padding, truncation. That text is the contract these functions exist to produce.
• ExpandInternal (32) — suppressed. Two over the threshold; one scan dispatching on tag kind.

Each suppression carries its own reason rather than a boilerplate one.

Test

• The relaxation is scoped, not blanket: with its NOLINT removed a production file reports the finding again (1), restored it reports none, and test files and mixed test+production invocations report none. A production file still gets fully linted (149 findings from other checks).
• bazel test --config=clang //mbo/mope/...:all //mbo/diff:all — 12/12 pass. The --set refactor is covered by //mbo/mope/tests/args, including the nested section:enabled,:config_start=25 path that exercises the extracted branch.
• pre-commit run -a green.