Beagle SCM

DIFF-005: be diff:<file> emits enormous (whole-file) context

be diff:dog/INDEX.md returns 110 lines for a 2-line edit — a single hunk @@ -1,106 +1,106 @@ spanning the ENTIRE file (2 added, 2 removed, 105 context). For a tiny change in a large doc this is "enormous context for no obvious reason." Likely DIFF-003's landed file-scoped full-file diff showing for a single-file diff:; research whether that is intended/acceptable or should be bounded. Evidence: evidence. Method: Issues.

Input

Context

• be diff:dog/INDEX.md (106-line file, +1 row edited) → 110 output lines, ONE hunk @@ -1,106 +1,106 @@, 2 + / 2 - / 105 context. The whole file is emitted as context for a 2-line change. Full capture + repro: evidence.
• DIFF-003 ("file-scoped diff: whole file, changes hilited") landed eaa75793, so a single-file diff:<path> plausibly renders the WHOLE file BY DESIGN. But the maintainer didn't expect/recognize it ("no idea why") — so either it's the intended file-scoped mode (and needs documenting / a bound for large files), or a context-sizing regression.

Goals

• Determine: is the full-file context for be diff:<file> INTENDED (DIFF-003 file-scoped) or a context-sizing bug? Cite the code path.
• Decide the right UX: if intended, bound it for large files and/or offer a git-style ~N-line-context mode, and DOCUMENT it so it stops surprising; if a bug, fix the context window.

Constraints

• Don't regress DIFF-003's deliberate full-file-with-highlight view where that is wanted (e.g. the pager); the concern is the plain be diff:<file> text output.
• Behaviour change to diff context gates for review; add a test pinning the chosen context size for a small-edit-in-large-file case.

WIP

Design decisions (scan done 2026-06-16 — INTENDED, not a bug)

• PREMISE CORRECTED (bare-be diff re-investigation): the bare aggregate be diff: is ALREADY bounded — every hunk is 3-line context (GRAFDiffWtTree→GRAFDiffWtFile(full=NO)→WEAVEEmitDiff). The whole-file inflation is ONLY the single-file diff:<file> projector (full=YES). So this ticket stays a single-file-projector concern; do NOT add bare-diff handling, and drop the "bare diff is whole-file" framing.
• The re-investigation surfaced two ADJACENT defects, now their own tickets: DIFF-006 (cosmetic: per-hunk --- a//+++ b/ repeated, not coalesced per file — the "62 headers vs 17 files" artifact, dog/HUNK.c:1242) and DIFF-007 (REAL bug: diff:<largefile> silently EMPTY — WEAVEEmitFull's 65536-token outtoks carve overflows on big files, NOROOM swallowed; WEAVE.c:1030/1099). Note: option-a here (bound large files → WEAVEEmitDiff) would INCIDENTALLY mask DIFF-007, a strong extra argument for it.
• VERDICT: the whole-file context is DELIBERATE DIFF-003 design, hard-wired full=YES for single-file diff: shapes. Decision site GRAFDiff2Layer (graf/DIFFREF.c:163): if (full) WEAVEEmitFull (no windowing, WEAVE.c:1015) else WEAVEEmitDiff (3-line WEAVE_CTX_LINES, WEAVE.c:802/850). Intent documented verbatim at DIFFREF.c:108-117 + GRAF.h:315-317. No context knob.
• Only the single-file diff:<file> projector is unbounded (GRAF.exe.c:512 wt-vs-base, :502 ref, :438 range). Tree-scope (DIFFREF.c:377) and the weave verb (GRAF.exe.c:567) ALREADY use the bounded 3-line path — so the machinery for git-style context exists.
• UX options (maintainer call — see Outcome): (a) BOUND LARGE FILES — at DIFFREF.c:163, when full && line_count > N (~200) demote to WEAVEEmitDiff; small files keep the full read (one-site, no API churn — worker's preferred). (b) SPLIT pager-vs-text — tty/HUNKMode → full, non-tty → 3-line (touches the 3 dispatch sites). (c) KEEP + DOCUMENT only (it self-documents already).

TODOs

• Scanned: full-file is DIFF-003-intended; window site DIFFREF.c:163; only single-file diff:<file> unbounded; tree/weave already 3-line.
• BLOCKED-ON-USER: maintainer picks UX (a) bound-large-files / (b) pager-vs-text / (c) document-only; then implement at the if (full) dispatch (now graf/DIFFREF.c:188-201 after drift) or just document.
• Test: small edit in a >N-line file → bounded context (alongside WEAVE2/01).
• Re-verified 2026-07-07 (ticket audit): still unbounded — GRAFDiff2Layer hard-wires WEAVEEmitFull on full (DIFFREF.c:188-196), no context knob; no landing commit in be log.

Blockers and bummers

• May be working-as-DIFF-003-intended — confirm before "fixing"; the deliverable might be a bound + docs, not a behaviour reversal.

Outcome

• Scan complete: INTENDED DIFF-003 behaviour (not a bug); window site graf/DIFFREF.c:163 (2026-07-07: drifted to :188-196). STILL awaiting the maintainer UX ruling (bound-large-files / pager-vs-text / document-only), then a one-site implementation. OPEN, blocked-on-user.