fix: encoding round-trip write side in skills_hub.py (ensure_ascii=False without encoding)

[AlexFucuson9]

Problem

3 write_text() calls in tools/skills_hub.py produce UTF-8 content via ensure_ascii=False but omit the encoding="utf-8" parameter. On Windows, write_text() defaults to the system codepage (cp1252), which corrupts non-ASCII characters in skill names, descriptions, and metadata.

PR #64822 already fixes the read side (read_text() → read_text(encoding="utf-8")). This PR fixes the write side to complete the encoding round-trip.

Why This Matters

Without this fix, the write path produces cp1252-encoded JSON on Windows, while PR #64822 changes the read path to expect UTF-8. This mismatch causes UnicodeDecodeError or silent data corruption when reading back cached skill data with non-ASCII content.

Files Changed

Line	Function	Call
1147	_write_index_cache()	cache_file.write_text(json.dumps(..., ensure_ascii=False))
3366	_write_stale_index_cache()	cache_file.write_text(json.dumps(..., ensure_ascii=False, default=str))
3406	SkillLockStore.save()	self.path.write_text(json.dumps(..., ensure_ascii=False) + "\\n")

[teknium1]

Thanks for covering the write-side calls: all three target writes are still unencoded on current main, so the diagnosis is valid.

Problems

• This cannot complete the claimed round trip by itself. The paired current-main readers at tools/skills_hub.py:1137, tools/skills_hub.py:3348, and tools/skills_hub.py:3402 still use read_text() without encoding="utf-8". PR fix: encoding round-trip for skills_hub/skills_sync cache and lock reads (supersedes #62667) #64822 contains corresponding reader changes but remains open, so merging this PR alone would write UTF-8 and then decode via the Windows locale.
• Please add a non-ASCII persistence regression. The existing TestHubLockFile.test_save_creates_parent_dir at tests/tools/test_skills_hub.py:1334 verifies only file creation.

Suggested changes

• Fold the three paired reader encodings into this change.
• Add a cache/lock write-read round-trip test using non-ASCII metadata.

Automated hermes-sweeper review.

[teknium1]

Merged via PR #71078 — your commit was cherry-picked onto current main with your authorship preserved in git log (rebase merge). Your 5-PR series (#50655 was authored under your earlier account, #54241, #56385, #66856, #65440) formed the backbone of the class-wide close-out: 68 of the 139 bare sites came from your commits, and the campaign's structure followed your directory-by-directory split. The remaining 71 sites were swept on top and a CI linter rule now prevents regressions. Thanks for the sustained, methodical work on this class.