Summary

Python's open() / Path.write_text() / Path.read_text() default to the
system locale encoding on Windows (CP1252 / CP932 / CP950), causing
UnicodeEncodeError when graph labels contain CJK characters or emojis.
This breaks file writes after LLM extraction is complete, discarding the
paid Claude API work from pass 3.

Also forces LF line endings for git hook installation, extending the fix
direction of commit 210243f ("fix hook reinstall, CRLF labels") to the
remaining hook write sites.

Concrete failure modes

1. GRAPH_REPORT.md / wiki .md / Cypher .txt writes crash on CJK Windows
   Unlike graph.json (which benefits from json.dumps's default
   ensure_ascii=True escaping), these files write raw CJK bytes. Under
   CP932 (JP) / CP950 (zh-TW) / CP949 (KR) locales, the final write step
   raises UnicodeEncodeError and destroys the extraction run.

2. Git hook CRLF breaks sh interpreter
   Path.write_text() without newline="\n" converts \n → \r\n on
   Windows, producing a #!/bin/sh\r shebang. Git Bash / WSL then reports
   bad interpreter: No such file or directory when the hook fires. This
   is the same class of bug as 210243f.

3. read_text() on UTF-8 files written by other platforms
   Reading graph.json / cache files / manifests created on Linux or
   macOS can fail on a Windows system with a non-UTF-8 default locale,
   even if the file itself is valid UTF-8.

Changes

All text file IO in graphify/ now explicitly passes encoding="utf-8":

No logic changes, no function signature changes, no behavior changes.
Binary-mode IO ("rb" / "wb") is untouched. newline="\n" is only
applied to git hook script writes — markdown / JSON outputs keep default
newline handling to match existing file conventions.

Tests

New file: tests/test_encoding_roundtrip.py (5 tests):

• test_write_read_text_cjk_roundtrip_primitive — baseline IO pattern
  with CJK + emoji + ensure_ascii=False JSON payloads
• test_cache_save_load_uses_utf8 — exercises
  graphify.cache.save_cached / load_cached with a CJK label
• test_to_json_open_pattern_cjk — reproduces the fixed to_json
open() pattern with ensure_ascii=False to force UTF-8 bytes into
  the output
• test_hook_install_lf_only — asserts
  graphify.hooks._install_hook produces no \r\n on fresh install
• test_hook_append_lf_only — same assertion when appending to an
  existing hook

The hook CRLF tests were manually verified to fail on the un-patched
v4 HEAD on Windows (raw bytes contain \r\n) and pass after the
fix — so they function as real regression guards rather than
tautologies.

All 5 tests pass on Windows Python 3.10.

Out of scope (intentionally)

• os.chmod(0o755) in hook install is still a no-op on Windows — that's
  a separate cross-platform concern and not addressed here.
• json.dumps(..., ensure_ascii=False) for human-readable graph.json
  output is a product decision, not touched.
• Performance / parallelism of the extract loop is out of scope.

Related

• 210243f — "fix hook reinstall, CRLF labels" (prior fix in the same
  bug class; this PR sweeps the remaining sites)