fix(config): preserve non-ASCII paths in written JSON configs (#497)

[Chang-Jin-Lee]

Linked issue

Closes #497 (remaining scope — see below)

What & why

json.dumps(existing, indent=2) defaults to ensure_ascii=True, so every config writer that round-trips a repo path or arbitrary pre-existing content through it serializes non-ASCII as literal \uXXXX escapes instead of native UTF-8. Six sites still had this:

• skills.py: install_platform_configs (.mcp.json cwd field), the shared _merge_hooks_into_settings (Claude settings.json), install_codex_hooks (Codex hooks.json), install_gemini_cli_hooks (Gemini settings.json), install_cursor_hooks (Cursor hooks.json)
• registry.py: Registry._save (registry.json)

Technically valid JSON either way, but as #497 documents, some MCP hosts / process launchers don't correctly decode \uXXXX when consuming these files directly.

One correction on scope: I believe #497's Bug 1 (json.dumps() abused for shell quoting, which is what actually produced the reported corrupted u57fa/u4e8e.../ desktop directory tree via double-encoding) is already fixed — the repo path is now substituted into hook scripts as a plain string, not JSON-encoded, so there's no longer a double-encode step for that specific flow. This PR closes out the remaining Bug 2 scope from the issue's "Affected locations" table: writing native UTF-8 in every config file this tool touches, which is still worth doing defensively (and is what a couple of those table rows asked for independent of Bug 1).

How it was tested

uv run pytest tests/ --tb=short -q   # 1967 passed, 5 skipped, 2 xpassed (pre-existing, unrelated)
uv run ruff check code_review_graph/ tests/test_skills.py tests/test_registry.py   # clean on touched files
uv run mypy code_review_graph/skills.py code_review_graph/registry.py --ignore-missing-imports --no-strict-optional   # no issues

Added one regression test per site (6 total, in TestNonAsciiConfigPreservation in test_skills.py and TestRegistryNonAscii in test_registry.py): seed non-ASCII content (a repo path or a pre-existing custom field), run the writer, assert the raw written file contains the literal UTF-8 characters and no \u escapes. Written first and watched fail (RED — every one failed with \uXXXX in the output) before adding ensure_ascii=False (GREEN).

Checklist

• Tests added for new functionality
• All tests pass: uv run pytest tests/ --tb=short -q
• Linting passes: uv run ruff check code_review_graph/
• Type checking passes: uv run mypy code_review_graph/ --ignore-missing-imports --no-strict-optional
• Lines are at most 100 characters
• Docs updated where behavior changed — N/A, internal serialization detail, no user-facing config/CLI surface changed