[mvanhorn]

Summary

rtk diff FILE1 FILE2 was printing both files in full, joined by a lone --- with no markers or hunk headers, for small changes instead of a real diff. The cause is the never_worse guard: it was handed a whole-file concatenation as its baseline, so for small files the correctly-condensed render had more tokens than that raw concatenation, lost the comparison, and the whole-file dump was shown. Because agents lean on rtk diff as an equality check, output that looks nothing like a diff is actively misleading.

This derives the never_worse baseline from a genuine diff instead, so the result is always a valid diff, either the condensed render or a classic normal-diff fallback with real NcN hunk headers, and never a whole-file dump. The identical-files short-circuit (exit 0) and the exit-1-on-difference contract are unchanged.

Test plan

• cargo fmt --all && cargo clippy --all-targets && cargo test
• Manual testing: rtk diff on a one-line change now prints a 1c1 hunk header with the changed lines instead of both files

Closes #2811

AI was used for assistance.

[KuSh]

Requesting changes — one of them blocks CI — but I want to lead with this: the core idea here is the most valuable one in the whole rtk diff PR pool, and it is under-sold by its own title.

Replacing content1 + "\n---\n" + content2 with a real classic diff doesn't just improve the fallback; it fixes the reason several other diff fixes are invisible to users. Because the old baseline is a dump of both files, never_worse prefers it constantly, so on small inputs rtk diff prints two blobs and the carefully-rendered output is thrown away. #3519's new whitespace message hits exactly this. Your PR is the structural fix underneath it.

Verification I did: built the head, fuzzed format_classic_diff over 900 random file pairs, and checked every emitted fallback with patch — all applied cleanly and reproduced file2 exactly, including tail-add, tail-delete, dissimilar-replacement, and the overlapping 1c1/1a2 anchor cases. No panics, no underflows, no non-advancing loops. That formatter is solid.

Four things to fix, three inline. The fourth: scripts/test-all.sh:532 asserts assert_contains ... "added", which never matches the new output (grep -c added = 0). It was already red on base, but this PR is the fix for that path and should update the assertion.

Also worth a look while you're here: docs/usage/FEATURES.md:249 still claims ~60% savings for rtk diff. On a 200-line file with 3 modified lines your output is 315 bytes against GNU diff's 315 — 0% reduction. The condensed render only wins past roughly 15 modified lines, because of its 2-line header. That number was wrong before this PR too, but the PR makes it easy to state the real one.

[mvanhorn]

All four are in as of bed658d, and thanks for the fuzzing writeup on the formatter, that was more verification than the PR had.

• The needless_borrow: dropped the & at the format_diff_changes call site. You were right that the PR body's clippy claim was wrong. cargo clippy --all-targets is clean now.
• The timer.track baseline: this was the sharper catch. raw had become the synthesized diff, which is empty for identical files, so gain reporting collapsed to ~0%. The baseline is now the original both-files dump again, kept separate from the fallback that feeds select_file_diff_output.
• The cfg(test) gate: render_file_diff is gone, folded into render_diff, and the tests call the real production path.
• scripts/test-all.sh:532: the "added" assertion is replaced with one matching the classic output and checking the exit code.

On FEATURES.md, I took your measurement rather than picking a new round number: it now says the saving is variable, 0% on small diffs, with the condensed render winning past roughly 15 modified lines.

Local run: clippy clean, 2438 tests passing.

[KuSh]

All four points from the previous review are fixed — threads resolved.

I rebased the branch onto develop and pushed it, since it had gone conflicting in the meantime. The only conflict was docs/usage/FEATURES.md: develop removed every **Economies :** line from that file wholesale in eb3f814, so your rewrite of the rtk diff one collided with a deletion. I dropped that hunk and trimmed the matching bullet from the commit message. No code was changed — scripts/test-all.sh and src/cmds/git/diff_cmd.rs are exactly as you pushed them.

Verified locally

• cargo fmt --all --check, cargo clippy --all-targets (with warnings = "deny"), cargo test --all — all clean, 2858 tests passing.

• The new test-all.sh assertion, run verbatim: passes.

• Recorded metrics compared against a develop build on identical fixtures and the same RTK_DB_PATH:

  case	develop	this branch
  identical files	[ok] Files are identical, 46.2%	unchanged
  one line changed	dumps both files, 0.0%	2c2 / < beta / --- / > betaX, 53.8%

• I fuzzed format_classic_diff over 900 generated file pairs (lengths 0–12 on both sides, so tail-append, tail-delete, wholesale-replacement and empty-file shapes all occur). 885 produced a non-empty patch; all 885 apply cleanly with patch and reproduce the second file byte-for-byte, and the 15 identical pairs correctly produced an empty patch. No panics, no underflows, no non-advancing loops.

Nice work — this is the change that makes a small edit readable instead of a dump of both files.

Two notes for the record, neither blocking

1. never_worse now guards the printed output against the classic diff, while timer.track records against the both-files dump — two different strings, where before they were one. So rtk gain can show a small negative for rtk diff on wholesale replacements (I measured −5% on rtk diff Cargo.toml LICENSE, where the old code clamped to 0%). It is arguably the honest number, since a faithful diff of two unrelated files genuinely isn't compressible. Flagging it so it isn't a surprise later.

2. The fuzzing also showed that only 289 of the 885 patches match what GNU diff emits. The rest are valid and exact but not minimal: compute_diff compares positionally, so a single inserted line shifts everything after it and cascades into a run of bogus NcN pairs. Across the corpus that makes the output about 40% larger than GNU's. That is pre-existing and well outside this PR's scope — your change is strictly better than what it replaces in exactly those cases — but it does bound how good the fallback can look until the underlying comparison is smarter.