fix(skills): handle bytes content in bundle_content_hash

[venyon2k]

Problem

bundle_content_hash() calls .encode("utf-8") on every file entry in SkillBundle.files, but the type annotation explicitly declares files: Dict[str, Union[str, bytes]].

When a SkillBundle is constructed by LocalSource.fetch() — which reads skill files from disk via f.read_bytes() — the files dict contains bytes values. Calling .encode() on a bytes object raises AttributeError.

Fix

Check isinstance(content, bytes) before encoding. Bytes content is fed directly into the SHA-256 hasher; string content is UTF-8 encoded as before.

Evidence

• SkillBundle annotation: files: Dict[str, Union[str, bytes]] (line 81)
• LocalSource.fetch() reads files as bytes: files[rel_path] = f.read_bytes() (line 2396 in current main)
• content_hash() in tools/skills_guard.py already handles both paths — it calls f.read_bytes() directly, the in-memory equivalent should match

Verification

• Existing test test_bundle_content_hash_matches_installed_content_hash passes (it uses string-only bundles)
• A test with a binary file in the bundle would demonstrate the crash

Co-authored-by: Jason Kuo jason@yechiu.com.cn

[alt-glitch]

Likely duplicate of #19120 (and #19081, #19079) — same fix for bundle_content_hash() calling .encode() on bytes values in SkillBundle.files.

[alt-glitch]

Likely duplicate of #19120 (and #19081, #19079) — same fix for bundle_content_hash() calling .encode() on bytes values in SkillBundle.files.

[teknium1]

The same bytes fix landed via #19575 (salvage of @teknium1's #19328, which was itself the canonical consolidation of five concurrent PRs — including yours). The logic you proposed is identical; thanks for catching and fixing this bug. Both contributors who implemented the same fix get credit via the commit's AUTHOR_MAP and this acknowledgement. Closing as duplicate.