mbo/diff roadmap

Gap analysis vs GNU diffutils and git's xdiff, 2026-07-04. The tooling is
functionally complete for its purpose (golden testing plus human review):
three engines (myers default, naive, direct) x four output formats
(unified, context, normal, side-by-side), POSIX exit codes, full
CLI/bzl parity, and a self-auditing engine x format test matrix
(diff_cli_sh_test.sh) that fails by name on any untested combination.
Remove items when done.

Easy

• -q/--brief: report only whether files differ (the exit code
  contract already exists; this just suppresses the output).
• --label/-L: override the header file names beyond
  --file_header_use.
• stdin via -: treat the file name - as standard input.
• Binary detection: print Binary files X and Y differ instead of
  diffing lines containing NUL bytes (today control bytes pass through).

Medium

• ed-script (diff -e) and RCS output formats: the last two classic
  formats. The per-chunk emitter dispatch in internal/output.cc takes
  new formats without structural work; the test matrix forces expected
  outputs per engine.
• Hunk sliding (git's indent heuristic): cosmetic post-pass moving
  hunk boundaries to visually pleasing positions; changes golden outputs.

Hard / needs design

• Patience / histogram engine variants: alternative Algorithm
  values on top of the token infrastructure; useful where unique-line
  anchoring reads better than minimal scripts.
• Wide-glyph / tab aware side-by-side columns: column arithmetic by
  display width instead of bytes (GNU expands tabs; wcwidth for CJK).

Non-goals (for the record)

• Recursive directory diff (-r): this is a two-file differ.
• GNU -y tab-stop padding byte-parity: we implement the well-defined
  --expand-tabs layout instead (local BSD and GNU diff differ here anyway).
• CLI flags for max_diff_chunk_length (naive-only safety valve) and
  time_format: intentionally library-only.
• Intra-line word/char highlighting: delta/difftastic territory.