Run clang-tidy 21 on pull-request changed lines

[matthargett]

What

Run the repository's existing .clang-tidy on pull requests — but only on the lines the PR changes.

The root .clang-tidy is already a maintained, strict config (bugprone-*, cert-*, clang-analyzer-*, ... with WarningsAsErrors), but no workflow runs it today, so the null-deref / narrowing-conversion / use-after-move class of bug it catches is currently caught only by a reviewer reading the diff.

How

Mirrors the diff-scoping already used by the git-clang-format gate in coding_guidelines.yml:

• permissions: contents: read, fetch-depth: 0, diff computed from ${{ github.event.pull_request.base.sha }}..HEAD.
• clang-tidy-diff.py turns the changed-line ranges into -line-filter, so a PR is only flagged for issues it introduces on the lines it touches — pre-existing issues on untouched lines stay silent.
• compile_commands.json comes from configuring the default linux iwasm build (interp + AOT runtime, no LLVM required); the database is emitted at configure time, so no build step is needed.
• Only changed sources that are in compile_commands.json are analyzed, each with its real compile flags. A changed file not in the default build — or a header, which has no translation unit of its own — is skipped rather than analyzed without context (which would otherwise produce false failures from missing includes or the wrong #ifdef branch).
• A PR that changes no C/C++ source skips install / configure / lint entirely, so docs-, test- and CI-only PRs cost almost no Action minutes while the job still reports success.

clang-tidy version

Runs clang-tidy 21 on ubuntu-24.04, installed from apt.llvm.org (a small ~50 MB package, not a full LLVM toolchain download). clang-tidy 21 matches the LLVM shipped in Xcode 26/27 and the clang-format-21 gate proposed in #4992, so lint results are consistent whether a contributor builds/auto-formats on macOS/Xcode or on another LLVM-built platform.

Curated .clang-tidy

Moving from clang-tidy 14 to 21 surfaces newer checks. Five are noise/style for this C codebase rather than bug-catchers, so they are disabled (on a full-tree scan they otherwise dominate the output with no correctness signal):

check	why disabled
misc-include-cleaner	IWYU-style; ~1750 findings, unusable noise
misc-header-include-cycle	structural noise
modernize-macro-to-enum	opinionated for a C codebase
readability-inconsistent-declaration-parameter-name	cosmetic
bugprone-assignment-in-if-condition	flags WAMR's deliberate assignment-in-condition idiom

The bug-catchers stay on, including bugprone-narrowing-conversions, bugprone-sizeof-expression, bugprone-implicit-widening-of-multiplication-result, bugprone-multi-level-implicit-pointer-conversion, performance-*, clang-analyzer-* and cert-* — the classes that have historically caught real portability / correctness bugs on this project.

Residual on existing code (why the gate can land now)

A full-tree scan with this config reports roughly 600 diagnostics on the default-config translation units — dominated by bugprone-narrowing-conversions — concentrated in core/iwasm/common and core/iwasm/interpreter. None of these block contributors: the gate is diff-scoped, so they only surface when a PR actually edits those lines. The residual is cleaned up incrementally in per-subsystem follow-up PRs (ordered stack in #4983).

Fork / upstream safety

Same properties as coding_guidelines.yml: only contents: read, diff from the base SHA — so it behaves identically for in-repo and forked pull requests.

Notes

• Coverage is scoped to the default linux config's translation units; it can be broadened with additional configs.
• Trigger is pull_request only: the gate diffs against the PR base, so a manual (workflow_dispatch) run would have no meaningful base.

Verification

Verified end-to-end against real CI on a throwaway fork PR:

• a deliberate bugprone-narrowing-conversions on a changed line of an in-database translation unit fails the job — confirming both that configure-only still parses the linux sources and that the diff-scoped fail path works;
• a PR that changes no C/C++ source skips the install / configure / lint steps and reports success (this PR's own run);
• a changed file not in the compilation database is skipped rather than mis-analyzed (diff intersected with compile_commands.json).

[lum1n0us]

Before turning this workflow into an enforced check, should we run a full  clang-tidy  scan on the current codebase and clean up any existing issues first? Otherwise, contributors may end up having to deal with pre-existing diagnostics while making unrelated changes in their PRs.

[matthargett]

Before turning this workflow into an enforced check, should we run a full clang-tidy scan on the current codebase and clean up any existing issues first? Otherwise, contributors may end up having to deal with pre-existing diagnostics while making unrelated changes in their PRs.

I tested this locally, and it's a big change. We can split it up and do it incrementally by subdir/component/owner, or do one rule per PR, or a combination of the two, or just do the big bang. I think the big bang sets a bad precedent we wouldn't want other contributions to replicate, and I don't have a strong opinion on which axes to break it up on.

when updating PC-Lint and eslint/prettier versions at previous jobs, beyond a pass to update disable/false-postiive comments, I let the reformats happen incrementally as changes came in (hence the approach in this PR). that said, it may not be great optics or overhead for a public/OSS project, so I'm flexible :)