fix: read the baseline and git history the same under any locale

[giacope]

When the locale isn't UTF-8 (LANG/LC_ALL unset, as in many containers), Ruby tags file and pipe input as US-ASCII. Any non-ASCII byte then made hashira stop with an internal error (exit 70).

What broke

• --ratchet / --update-baseline with a baseline holding non-ASCII text, e.g. an em dash in an acceptance reason. kimera's hashira_baseline.json hits this:

  hashira: internal error — Encoding::InvalidByteSequenceError: "\xE2" on US-ASCII
    at .../json-3.0.2/lib/json/common.rb:302:in 'String#encode'
  

  Baseline#read did JSON.parse(File.read(@path)), and JSON could not transcode the US-ASCII-tagged text.
• Every run, when git's core.quotepath is off and the history names a non-ASCII file. GitLog#counts split the US-ASCII-tagged git log output and got ArgumentError: invalid byte sequence in US-ASCII.

Changes

• CI::Baseline reads and writes the baseline as UTF-8. The write was already correct under a plain non-UTF-8 locale. With default_internal also set (e.g. ruby -U), it transcoded UTF-8 to US-ASCII and raised UndefinedConversionError. Now it can't.
• GitLog reads git log as bytes (binmode: true), and Churn#hits compares paths byte for byte. Decoding git's output as UTF-8 would not be enough: under a US-ASCII locale Dir[] returns non-ASCII paths as BINARY, and end_with? raises on UTF-8 vs BINARY. Comparing bytes matches the names under either locale.
• spec/readme_spec.rb read README.md the same way and failed under that locale. It now reads it as UTF-8, so all of bin/ci passes with LANG unset.
• Other inputs checked and left alone: source files go through Prism.parse_file, which reads bytes and defaults to UTF-8. The JSON and text reports print correctly under that locale. hashira has no config file.

Tests

• A new with_default_encodings spec helper swaps Encoding.default_external/default_internal and restores both in an ensure.
• New specs:
  • A baseline with non-ASCII text loads under US-ASCII.
  • A baseline write under US-ASCII external + UTF-8 internal writes the UTF-8 bytes.
  • A repo with core.quotepath=false and a file named café.rb gets churn 1 under both UTF-8 and US-ASCII.
• All three fail on main with the errors above. Each change was also reverted alone (.b, the write encoding, the read encoding), and each revert fails its own spec.
• bin/ci passes with LANG=C.UTF-8 and with LANG/LC_ALL unset: rubocop, rubocop-kata doctor, 502 examples, the gate, and Ratchet OK: 16 edges, 0 findings, unchanged.
• kimera ci --since origin/main: 5 mutants on changed lines, all killed.
• End to end, on a throwaway project with lib/café/menú.rb and core.quotepath=false under an unset LANG: --ratchet, --update-baseline and a plain run all exit 0 (before: 70), and churn counts café/menú.rb.

Not in this PR

With git's default core.quotepath=true, git log quotes non-ASCII paths ("lib/caf\303\251.rb"). Churn then never matches those files and reports 0 under any locale. Passing -c core.quotePath=false in GitLog::LOG would fix it. It's a separate bug, so I left it out.

🤖 Generated with Claude Code

https://claude.ai/code/session_01PcDfzLuRgjoSnfPLrTaA4h

---

Generated by Claude Code