[kylehgc]

Adopts upstream rtk-ai/rtk#3268 by @iliaal. Closes #100.

What this fixes

Two defects in src/cmds/git/diff_cmd.rs, both reachable from a plain rtk diff a b (and from bare diff, which the hook rewrites into this path):

1. False [ok] Files are identical with exit 0 for files whose only difference is CRLF-vs-LF or a missing final newline — identity was decided from str::lines(), which normalizes both away. A false identity silently passes any verification gate built on diff.
2. Positional line alignment: one inserted line desynchronized every later comparison, the whole tail rendered as changed, and never_worse then dumped both files concatenated (403 lines for a 1-line insertion).

Commits

Commit	Author	Kind
ce07aff fix(diff): compare bytes before claiming two files are identical	Ilia Alshanetsky	cherry-pick of upstream c1d47d6
54cb3a1 fix(diff): align by LCS instead of comparing line N to line N	Ilia Alshanetsky	cherry-pick of upstream d4411cc
c7d29b0 test(diff): cover the LCS_CELL_CAP wholesale-replacement branch	kylehgc	fork-authored amendment (review round)

At the cherry-pick tip (54cb3a1), git diff pr-3268 HEAD -- src/cmds/git/diff_cmd.rs was empty — the adopted code is byte-identical to the upstream PR branch. The amendment adds one test on top (28 insertions, test module only) and does not touch production code.

Repro (real binaries, Windows)

Fixtures: base.txt = 200 lines; base_crlf.txt = same text, lines 50–74 CRLF; inserted.txt = base with one line inserted at the top.

$ cmp base.txt base_crlf.txt
base.txt base_crlf.txt differ: char 1041, line 50   (exit 1)

# BEFORE (installed rtk 0.42.4, pre-fix develop)
$ rtk diff base.txt base_crlf.txt
[ok] Files are identical
exit: 0

$ rtk diff base.txt inserted.txt | wc -l
403        # never_worse dumped both files concatenated

# AFTER (target/x86_64-pc-windows-msvc/release/rtk.exe from this branch)
$ rtk diff base.txt base_crlf.txt
base.txt → base_crlf.txt
   differs, text matches (line endings: 0 CRLF vs 25 CRLF)
exit: 1

$ rtk diff base.txt inserted.txt
base.txt → inserted.txt
   +1 added, -0 removed, ~0 modified

+   1 INSERTED
exit: 1

$ rtk diff lf.txt lf.txt          # truly identical files unchanged
[ok] Files are identical
exit: 0

Note: for very small differing files the never_worse guard prints the raw concatenated content instead of the "differs, text matches" message (condensed is longer than raw at that size). The exit code — the part scripts rely on — is correct in every case. Same caveat upstream rtk-ai#3471 documents.

Quality gate

cargo fmt --all (no changes) && cargo clippy --all-targets (zero warnings) && cargo test --all (2691 passed in the main suite, 0 failed across all suites), via the x64 host toolchain (scripts/win-dev-env.ps1).

No pre-existing fork tests were modified; the two cherry-picks add 11 new unit tests and the amendment adds 1 more (LCS_CELL_CAP branch).

Deliberately not adopted

• Upstream #3471 (byte-accurate compare + --ignore-whitespace flag) and #3519 (byte-guard only) patch the same site. fix(diff): stop reporting differing files as identical, and align by LCS rtk-ai/rtk#3268 is the most complete — it is the only one that also fixes the positional-alignment defect — so it was adopted and the others are referenced here per the ticket. The --ignore-whitespace convenience flag from fix(diff): compare files byte-accurately instead of by normalized lines rtk-ai/rtk#3471 is a feature, not a bug fix, and was not adopted.

[kylehgc]

Review round (pre-merge, two-axis /code-review vs develop)

Standards axis — no hard violations. Two judgement calls, both deliberately left as-is for upstream byte-parity with rtk-ai#3268:

• Per-changed-line String allocation in the Op edit script (rust-patterns anti-pattern "clone of large strings"); a borrowed Op<'a> would avoid it. Cold path (only allocates for changed lines), not worth deviating from upstream.
• No count_tokens ≥60% savings test in this module (cli-testing checklist). Pre-existing gap for the whole file-diff module; the savings baseline for a two-file compare is ill-defined. Left for upstream to define.

Spec axis — algorithm verified end to end (LCS backtracking hand-traced, prefix/suffix trim guards checked for overlap, CRLF counting matches the spec message, exit-1 propagation traced to std::process::exit). Two findings:

• Provenance references: already carried in this PR body (adoption of fix(diff): stop reporting differing files as identical, and align by LCS rtk-ai/rtk#3268, duplicates fix(diff): compare files byte-accurately instead of by normalized lines rtk-ai/rtk#3471/fix(diff): don't report byte-different files as identical rtk-ai/rtk#3519 referenced). No action.
• Fixed: the LCS_CELL_CAP fallback branch had no test — the existing wholesale-replacement test uses 4 cells and never reaches the cap. Addressed in fork-authored amendment c7d29b0 (2001×2001 fixture, > 4M cells).

Quality gate re-run after the amendment: fmt clean, clippy zero warnings, cargo test --all green (2692 passing in main suite).