zhtw-mcp: Skill for Claude Code

Validating a zhtw-mcp change

One gate, and it is the thing to run:

make check       # cargo test, clippy on three feature shapes, the formatters,
                 # the ruleset lint, the hook suite, shellcheck
make check-size  # the release binary has to stay under 20 MiB

Python 3 is a build requirement, not just a test requirement: make regenerates src/engine/s2t_data.rs from the pinned OpenCC dictionaries before it builds.

What skips, and why that matters

Every formatter lane skips when its tool is missing rather than failing, and that is every tool scripts/indent.sh drives: commentflow, cargo, black, shfmt and python3, plus shellcheck in the Makefile. That is right on a laptop and wrong on a runner, so CI installs them on the Linux leg and sets ZHTW_REQUIRE_TOOLS=1 there, which turns a skip into a hard failure. The macOS leg still skips, and the verdict line names what it skipped. A green local run is therefore weaker evidence than a green CI run, and so is what the local gate leaves out: CI also holds a pull request's own commit messages to the rules, runs cargo audit against Cargo.lock, builds the browser extension, and runs the whole suite on macOS and Windows. Two of the three Windows-only regressions this project has shipped were found by that last one and by nothing else.

The indent gate is the one that surprises people

scripts/indent.sh --check copies the tree, runs the whole formatter chain over the copy, and diffs: comment reflow with commentflow, then cargo fmt, black, shfmt, and the assets/ruleset.json normalization that scripts/check-ruleset.py owns. Checking the composition rather than each tool is not a flourish. commentflow puts a blank line before a comment inside a method chain and cargo fmt takes it straight back out, so commentflow --check alone can never be satisfied on Rust. make indent runs the same script with --write, so the fix for a failure is always that one command.