Record the sync point a publish creates

[rhiever]

goodeye skills publish moved the registry without ever touching the local sync index. A directory publish set both drift signals at once: the uploaded tree left the recorded hashes on superseded content while the version advanced, which classifies as a conflict. Nothing could clear it, because a pull refuses a dirty tree and a push sends a token the registry has already replaced, so the mirror stayed stuck until a forced pull discarded the local edits it was protecting.

What changes

A directory publish records the whole sync point, found by the published directory's own path. The tree was uploaded, so the registry holds what the directory holds whatever version the index was sitting at, which also recovers a mirror already stranded.

A body-only publish reads no directory, so it records only what it knows, only on mirrors sitting at the version it replaced, and writes the new body to disk only when nothing unsaved is there to lose. A mirror holding unsaved edits keeps them and reads as ordinary drift a push resolves.

Publishing a directory also cleared every sibling-file label and execute bit. Both live only in the registry, and a snapshot that omits one clears the stored value, so a tracked mirror's recorded labels and execute bits now ride along with the upload the way a push already sends them. For an unchanged file the recorded execute bit wins, which is the rule build_files_payload already applies on a push, so publish and push cannot disagree about the same directory. build_files_payload itself is untouched.

Verification

• Full suite: 1125 passed. ruff, ruff format and pyright all clean.
• The new tests were run against the unfixed source: 7 of 13 fail. The 6 that pass unfixed are the "leaves the index alone" guards, which hold trivially when nothing writes the index, so they guard against over-reach rather than reproduce the bug.
• The best-effort contract is covered in both directions: an unreadable index or an unwritable mirror warns without failing a publish that already landed on the registry.

Not in scope

• A publish of an untracked directory still clears labels. The server's save path snapshots while its patch path carries labels forward; making the two consistent is a server-side change.
• Prompt packs that pipe a body-only rewrite through publish - would be better served by put-file.
• The public docs do not yet describe the new local side effects of a publish. Tracked as a docs follow-up.

Release

Version bumped to 0.25.3. This release carries only the bug fix: the single-path file commands shipped in 0.25.2. Not tagged yet. The changelog entry lives in the server repo and ships as a separate PR.

🤖 Generated with Claude Code