[foxnne]

Part of fizzyedit/plans plans/DIFF_PLAN.md step 1 (2.1-2.2, 3).

What

Std-only core.diff (core/diff.zig, core/diff/{Text,unified,tests}.zig):

• Text: reusable scratch; byte prefix/suffix trim with @Vector(32, u8), line interning (Wyhash, byte-exact on match), histogram diff (occurrence cap 64) with an explicit stack, Myers (linear space, max_cost cap, then the range is replaced) when there is no anchor, hunk sliding to blank-line/low-indent boundaries. Hunks carry lines and bytes; Text.edit gives a byte-range Edit; Text.range converts to protocol positions through the owner's byte-to-position callback.
• unified.write: standard @@ -a,l +b,l @@ with context, \ No newline at end of file, and a max_bytes cap that drops whole trailing hunk groups and says ... N more hunks.
• bench-diff + tests (round trip over random edits on text/CRLF/no final newline/binary, scratch reuse allocates nothing, unified fixtures and cap), wired into build/app.zig.

Not in this step (named seams): budget_ns partial/resume, word/char refinement and merge (step 2), cmdWrite untouched.

Bench (zig build bench-diff -Doptimize=ReleaseFast, 1.15 MB / 25k lines, Linux)

Case	Time	Budget
identical	0.10 ms	0.2
one line inserted	0.07 ms	2
1% scattered (250 hunks)	13.2 ms	15
every line changed	1.7 ms	60
two-line alphabet (Myers, capped)	16.3 ms	60

Prior-art notes in the plan (git histogram, JGit, imara-diff) were not re-fetched; the implementation follows git's histogram scoring from memory and diff-match-patch's bisect for Myers.

Verified

scripts/check-fmt.sh, zig build test, check-web, check-tests-reached pass. No SDK boundary change.

---

Generated by Claude Code