clang-tidy: Make it actually useful.

[michalhosna]

lint.sh runs nowhere in CI, and .clang-tidy sets Checks: *, which enables every vendor check shipped with clang-tidy.

Why main looks green

.clang-tidy carries AnalyzeTemporaryDtors, removed in clang-tidy 16. clang-tidy rejects the key, exits 0, and falls back to the default check set — so Checks: * has never actually applied. Measured on one TU: 1 warning under main's config, dozens once the key is dropped. lint.sh passing on main means nothing.

scripts/lint.sh on main is also run-clang-tidy -p build with no source filter, so much of what it walked was CPM/dependency code rather than ours.

What a real run finds

With the key dropped, over src/, test/, benchmark/ — roughly 18,000 unique warnings, dominated by checks written for other projects' coding standards:

group	share	note
llvmlibc-*, altera-*, fuchsia-*, zircon-*, …	~2/3	vendor checks for LLVM's libc, FPGA compilers, Fuchsia — none apply here
misc-include-cleaner, readability-identifier-length, readability-magic-numbers	most of the rest	style/opinion
bugprone-*, performance-*, clang-analyzer-*, concurrency-*, cert-*	~350	the part worth acting on

The config is the problem, not the code.

Proposal

• Replace Checks: * with an explicit list, dropping the vendor families outright.
• Triage the ~350 bugprone / performance / analyzer findings.
• Only then add a lint job to ci-pr.yml next to the existing format.sh --check.

Note on gating: WarningsAsErrors: '' means warnings alone don't fail the run, but run-clang-tidy still returns non-zero if any TU emits a clang-diagnostic-error. A gating job wants WarningsAsErrors set to the curated list, so warnings actually fail it.

_{Counts are approximate, from a full clang-tidy 19.1.7 run on Debian}

[afrind]

Oof, nice find.