fix: read the current markdown body in get

[sb123sb123]

pyrite get --format json now reads the current Markdown body and raw-byte SHA-256 for the selected entry, even when the index is stale. Passing that body to pyrite update --body-file leaves the file unchanged.

Scope: the file-backed refresh is opt-in and used by CLI get; REST, MCP, and other KBService reads keep their established normalized-body behavior.

• Fixes kb get returns a body missing a section; a get → update --body-file round trip deletes it #753.
• Regression starts with a stale index, checks the exact current body and hash, then feeds the body to update and verifies identical file bytes and an empty Git diff.

Validation:

• Ruff check, Ruff format, and git diff check pass on Windows.
• Focused Windows run: 145 passed; two path-separator assertions and 11 SQLite teardown errors are Windows-specific. The new round-trip regression and core-test selector pass.
• The first hosted run exposed the side effects of refreshing every KBService caller. The follow-up commit narrows the refresh to CLI get; hosted CI is rerunning on this head.

[markramm]

Automated review (Pyrite conductor, an AI agent) of head 034aeba. The maintainer decides; this is input only.

Thanks, @sb123sb123. The fix does stop the reported loss: with a stale index, the ### QC section now survives get → update --body-file, and the regression test fails without the change. Four things should change before this merges.

1. Use the one splitter. refactor: one frontmatter splitter, every reader uses it #744 has merged split_frontmatter (pyrite/storage/frontmatter.py) as the single frontmatter splitter. _body_from_entry_file adds a third rule (line.strip(" \t") == "---") that disagrees with it. For example, ---\nid: n1\ntitle: T\n--- # end\n\nIntro kept?\n\n---\n\nTail.\n returns Tail.\n, so a get → update round trip would delete "Intro kept?". Please rebase on dev and take the body span from split_frontmatter.
2. One meaning of "the entry's body". The file read is behind a CLI-only flag, so REST (endpoints/entries.py) and MCP (mcp_server.py) still return the index body, and the same round trip loses data there. ADR-0042 decision 9 says a single-entry read reads the file, for every surface. Please move the read into get_entry itself rather than behind read_current_body.
3. One read, one version. Take title, tags and the other fields from the same file read as body and content_hash. Today a stale index can return old frontmatter beside a fresh body and a hash of the new file.
4. Claim only what holds. update rewrites the whole file, so "leaves the file unchanged" is false for files with CRLF, a BOM, no final newline, YAML comments or extra trailing blank lines. Byte-identical body writes come with B6 (B6 write path: phases, acceptance and footprint (spike result) #730, apply()). Please narrow the changelog to "get returns the file's current body". Add round-trip tests built from hand-written bytes for those file shapes, not from to_markdown(). Assert body equality for get, and file equality only where it actually holds.

Smaller: an entry added since the last sync still gives NOT_FOUND. find_file scans the whole KB when the id is not the filename.