Beagle SCM

GET-021: a worktree silently diffs against an advanced cur it was never materialized to; be status must warn when stale

A worktree's be diff/be status baseline is its cur POINTER, but a selective be post (and an incremental be get) can advance cur past commits that changed files NOT in the worktree's working set — leaving those files stale on disk while cur's tree already has the new content. be diff then renders a SPURIOUS reverse-diff (disk-old vs cur-new) and be status flags mod/mis for files the user never touched, with NO hint the worktree is behind. A worktree should diff against the rev its files were actually PRODUCED FROM, and be status should WARN when that produced-from rev ≠ the branch tip. Method: Issues.

Input

Context

• Repro (this session, ~/beagle): sibling worktrees landed commits into the

  shared store (e.g. BLAME-006a fa09701f changed graf/BLAME.c, 849→882 lines), never touching ~/beagle's disk. A later SELECTIVE be post from ~/beagle (staging only 3 unrelated graf/{GRAF,LOG} files) rebased cur onto fa09701f → f6bea1c5, and a get ?#f6bea1c5 did not re-materialize the sibling-landed files. Result: ~/beagle/graf/BLAME.c on disk = the OLD 849-line file (12 fix-markers), cur f6bea1c5's tree = the NEW 882-line file (22 markers).

• Symptom: be diff shows graf/BLAME.c + graf/test/CMakeLists.txt as edits the

  user never made (a REVERSE-diff: disk behind cur), and be status lists them mod/mis — indistinguishable from real WIP. Hit twice this session (also the ~/todo/ABC-003 worktree). Repeatedly read as "the worktree is broken".

• LANDMINE: be put graf/BLAME.c + post from this state would COMMIT the stale

  old file and silently REVERT the sibling-landed fix.

Goals

• be diff (and be status) in a worktree baselines against the rev the working

  files were actually materialized from — not a cur pointer that was advanced past them — so stale-but-unedited files are NOT shown as user changes.

• When the worktree's cur ≠ the branch tip, be status PREPENDS a commit-difference

  block (before the per-file rows), one row per differing commit — post for AHEAD (local commits not in tip → would be posted), miss for BEHIND (tip commits not local → missing here) — mirroring how get/post/head report ahead/behind. Shape:

      status:
       post 49039f36  11:42  GET-020: …            ← ahead (local, unposted)
       miss 3dc19b0a  12:40  DIFF-002: …           ← behind (in tip, not here)
       10:29  mod ABC.md
       ? 1212 ok, 1 mod   (behind 1, ahead 1)
  

Constraints

• Don't regress the legitimate mod/mis reporting of real working-set edits.
• Prefer not to widen every post/get into a full re-checkout if a cheaper

  detect-and-warn (compare cur's tree-of-changed-paths vs disk) suffices; but a complete materialization on cur-advance is the alternative (cousin of GET-018 atomic checkout / GET-020 incomplete checkout, and GET-017 ref-advance).

WIP

Design decisions

• CHOSEN (2026-06-16, user): DETECT-and-show. be status renders the cur-vs-tip

  commit difference as post (ahead) / miss (behind) rows ABOVE the file rows — reuse the EXISTING ahead/behind enumeration, not a re-roll: be head's local ?br ahead/behind path (beagle/BE.cli.c:1538, via GRAFHead/GRAF.exe.c:535) already computes the two commit sets; status just needs to emit them as {verb=post|miss, uri=commit:?<sha>} hunk rows (clickable, like log: rows now are). NB: the help already claims "Bare be = status (current branch, ahead/behind, dirty)" (BE.cli.c:89) — so the intent is documented; only the per-commit LIST is unimplemented. This is a spec/impl gap as much as a feature.

• Materialization (the stale-FILE half) is secondary/optional here: showing the

  miss commits already warns the user to be get; fully re-checking-out on cur-advance is the heavier alternative (cousin of GET-018/GET-020).

TODOs

• Repro-first: a hermetic two-worktree store — wt-B lands a commit, wt-A is left

  behind; assert be status in wt-A prepends a miss <sha> … row (and a post row when wt-A is also ahead), and that the row count clears after be get.

• Wire be head's local ahead/behind commit sets (BE.cli.c:1538/GRAFHead)

  into be status output as post/miss commit:?<sha> hunk rows above the file rows; keep file-status reporting unchanged.

• (optional/secondary) re-materialize stale cur-delta files so be diff stops

  showing phantom reverse-diffs — or document that miss rows + be get cover it.

Blockers and bummers

• Needs the worktree's true materialization base tracked/derivable separately from

  the cur pointer (or a tree-vs-disk compare). Pairs with GET-017/GET-018.

Outcome

• LANDED ef8901fa (FF child of addd5a9c; 309/309; 6 files): be status PREPENDS the

  cur-vs-tip commit block — post rows (ahead) / miss rows (behind), newest-first, clickable commit:?<sha> — above the file rows, with (behind N, ahead M) in the summary; cleared after be get, up-to-date output byte-identical. Reused SNIFF's OWN DAG walk (no cross-process orchestration): GETStatusCommitDiff(cur,tip,…) + GETLocalBranchTip (empty branch → trunk ? REFS row, the fix that makes trunk worktrees work) in sniff/GET.c; status_emit_commit_diff in sniff/SNIFF.exe.c:386; miss verb palette slot in dog/ULOG.c. New beagle/test/be-status-ahead-behind.sh (RED→GREEN). 309/309. Awaiting gate. Msg: GET-021: be status shows ahead/behind commit rows. (This is the detect-and-show half; per-file stale-materialization is the secondary, left to a follow-up — the miss rows already warn the user to be get.)