Beagle SCM

DIFF-016: EXPECTED-based diff — green/red local, pale blue/orange patched-in, yellow conflicts

Per DIS-080: with fence markers retired, the diff view is where conflicts are seen and resolved. EXPECTED = the straightforward weave reading of base + all in-scope patch-ins (theirs layers, no WT layer). Green/red = wt compared to EXPECTED (local outstanding edits). Patched-in changes (EXPECTED vs base) render pale blue/orange — the other kind of in/rm, not salad/salmon. Tokens where both sides touch (conflicts) render yellow. All provenance comes from the weave — per-token origin ids — no separate diff3 pass.

Input

Context

• Today's diff shows one green/red axis vs base; a patched-in change is indistinguishable from a local edit, and a conflict was only visible as fences in the bytes.
• The weave already carries per-token source ids (commit hi64, WT layer 00000000005774ed); ours/theirs/both classification per token falls out of the fold.
• Multi-patch runs stack: EXPECTED folds ALL patch rows in scope, several runs compose in one weave.

Goals

1. Diff render classifies each changed token: theirs-only → pale blue (in) / pale orange (rm); wt-vs-EXPECTED → green/red; overlap (conflict) → yellow. Conflict paths cross-checked against ULOG-004 accessor.
2. Row metric: patched-in tokens weigh as edits — in/rm in the max(in,rm) row weight, total = max + eq ([inline-vs-split-metric-max] extends, C HUNK.c + JS bro.js in lockstep).
3. Theme: pale blue/orange + yellow SGRs in view/theme.js, plain-text fallbacks for goldens; quad/status colors stay red for con rows.

Constraints

• Weave-based only — no diff3 reimplementation, no marker parsing.
• Verify through the real UI path (pager pty), not TLV probes ([verify-through-real-ui-path]).

WIP

Design decisions

• EXPECTED reading = shared/expected.js expectedOf() — patch.js mergeApply MINUS weave.foldWt: weave.build per side, weave.merge, mergedLive.
• Provenance rides the EXPECTED bytes as a THIRD weave layer folded BETWEEN base and wt (ID_EXP), so ONE emitDiff spells side AND origin — no 2nd pass.
• emitDiff/emitFull emit the weave walk verbatim (verified: tokens 1:1, never split, windows contiguous) — a hunk record aligns to it by token text.
• Encoding = tok32 bit 26 (custom) means PATCHED IN. C HUNK.c IGNORES that bit (probed), so --plain stays byte-identical; side 3 was rejected (C plain reads it as EQ and collapses the +/- pair).
• CONFLICT needs no third bit: it is a RUN of consecutive changed tokens carrying BOTH bit-26 values — the same membership test mergedLive uses. EQ tokens break runs, so two edits on neighbouring lines never false-yellow.
• PLAIN-TEXT FALLBACK = NONE, deliberately. test/diff/18-patch-base asserts ^\+THREE$ on --plain after a patch and I may not edit it; provenance is a colour distinction, asserted on --color SGR digits (no ESC literal, so dash-safe). --plain output is unchanged for every provenance.
• Yellow is derived from the WEAVE (DIS-080: regions come from the weave, not rows). wtlog.conflicts() is read in diff.js's expectedCtx and cross-checked at TEST level (status con tally == the yellow path) — path-granular only.
• Colours (view/theme.js DIFF_WASH, 256-cube bg): local 194/157 + 224/217 (unchanged), patched-in 153/117 (blue) + 223/215 (orange), conflict 229/227 (yellow — 229 not 230, which is the banner band).

TODOs

• shared EXPECTED reading (weave minus WT layer) — same object post's fold commits
• token provenance → color class in the diff emit; theme entries
• metric: JS side done (bro.js classifyLines pIn/pRm + max(in,rm)+eq)
• HUNK.c twin — jab/ is READ-ONLY here; follow-up spelled out below
• tests: local-only, patch-only, mixed, conflict, two patch runs stacked

HUNK.c follow-up (the lockstep C twin, NOT done here)

• File beagle/dog/hunk/HUNK.c, fn bro_classify_lines (the twin of view/bro.js classifyLines) + bro_classify/lineKind.
• Rule: a token with tok32 bit 26 (custom) set is an EDIT — its bytes go to in_b/rm_b, NEVER eq_b; row weight stays max(in_b,rm_b), total max + eq_b ([inline-vs-split-metric-max]). Break out p_in/p_rm to mirror the JS tally.
• bro_cell_ansi needs the same wash families (THEME slots for pale blue/orange + yellow) and the run-overlap → yellow rule, or the C pager will keep painting patched-in tokens salad/salmon.
• Same file also carries DIFF-019 HIGH: HUNKu8sFeedLineBased fixed 64KB/4KB side buffers throw hunk.render: out full on big files and SILENTLY truncate regions past 4KB (bare u8bFeed, refusal dropped) — the one libdog pass should fix both while touching HUNK.c.

Blockers and bummers

• Depends on PATCH-025 (markerless mergedLive), which is NOT merged — its 3 files were imported into this worktree by orchestrator order. DIFF-016 cannot land before PATCH-025.
• RESOLVED 2026-07-30: the five PATCH-025 regolds (get/conf-state, get/confdrop, get/dirty, post/conf-state, status/conflict) ported byte-identical from //PATCH-025/test (gritzko-approved versions), all green + staged here. Sole remaining failure: post/patch-absorb (POST-035's row-based gate) — 320/321.
• 2026-07-30 base moved to ///be/#5c0d5c21: the wt's own markerless weave conflicted get.js/weave.js and token-glue left get.js unparseable (267 fails); resolved by hand — weave3 (now in shared/weave.js) renders via mergedLive returning {bytes, spans}, both callers adapted (get.js:1275 binding restored, shared/checkout.js spans-not-markers — checkout.js is NOW PART of this tree's staged set). Suite back to 315/321.
• NO diff golden anywhere needed a regold — the plain path is byte-identical.

Outcome

• Not committed. Suggested commit message: DIFF-016: EXPECTED-based diff colours
• Staged in //DIFF-016 (jab put): new shared/expected.js; edited views/diff/diff.js, view/bro.js, view/theme.js; imported (PATCH-025) shared/weave.js, verbs/patch/patch.js, verbs/get/get.js.
• Staged in //DIFF-016/test: new test/diff/{19-patch-provenance, 20-patch-conflict,21-patch-stacked}; imported PATCH-025 conflict goldens.
• Full suite 316/321 (the 5 imported PATCH-025 failures, no new ones); diff 20/20, patch 28/28, status 14/14.
• Pager verified through a pty-driven Pager.render() probe (tty.openpty, raw mode): all five washes reach the real frame bytes.