The mental model

Five commands over one tracked document:

   snapshot ──►  (you edit the file)  ──►  sync  ──►  (clean)
      │                   │                  │
      │                 diff                 └─► conflict ──► resolve ──► sync
      │            what have I changed?
      └─ start tracking                    status  (read-only, any time)

Three ideas explain almost everything.

Three revisions, one merge. Every tracked document has a base (B, the upstream revision you last synced against, snapshotted under .cedit/base/), a local copy (L, the file you edit in place), and an upstream (U, the new revision you hand to sync). sync aligns L against B to learn what you changed, aligns U against B to learn what upstream changed, and decides every block by crossing the two. The merged document is always U's structure with your texts spliced in.

Blocks are content-addressed, and fences count. A "block" is a paragraph, a heading, a table cell — or a code fence, a raw HTML block, or the YAML front matter. Its identity is a 16-hex-char Merkle hash of its canonical content, not its position or line number. That is why a rewritten fence is a first-class edit, why an upstream move of a block you edited costs nothing, and why rewrapping a paragraph at a different width changes nothing at all.

Nothing is ever silently clobbered. When upstream changes the very block you adapted, the working file keeps your text and the conflict is recorded with all three versions. The document then refuses to sync again until you settle it with resolve. There are no conflict markers in the file — ======= is a setext heading underline, so a marked-up Markdown file would stop parsing as the document it was; conflicts live in .cedit/manifest.json instead.