feat(diff): context/normal output formats, Myers algorithm as new default

[helly25]

Summary

Three related changes to mbo/diff, developed and verified together:

1. Output formats: unified | context | normal

• New DiffOptions::output_format / --format flag / format attr on the bazel diff_test rule.
• Context format (diff -c) and normal format (plain diff) reproduce GNU diff byte for byte (verified against the system diff on 18 case families incl. empty files, no-newline markers, multi-hunk inputs) and apply cleanly with patch(1).
• Rendering was extracted out of internal/chunk.cc into internal/output.{h,cc} (one emitter per format); Chunk now only accumulates and filters, and file headers are format aware (***/--- for context, none for normal).
• Bug fix: unified chunk headers for empty ranges (pure insertions/deletions, visible with --context=0) now reference the line preceding the gap (@@ -2,0 +3 @@ instead of @@ -3,0 +3 @@), matching GNU diff. Previously patch applied such hunks one line too late. Found by round-tripping every format through patch.

2. Myers diff algorithm — the new default

• impl/diff_myers.*: linear-space middle-snake divide and conquer ("An O(ND) Difference Algorithm and Its Variations", the algorithm behind GNU diff and git) over interned line tokens, so every ignore/strip/replace option applies unchanged. Produces minimal diffs; past a cost cap of max(64, √(L+R)) it takes a git-style furthest-reaching split to bound pathological inputs.

• Replays through the existing ChunkedDiff push interface: chunking, context handling, filters and all three output formats are shared.

• Benchmark (bazel run -c opt //mbo/diff:diff_benchmark, 10k-line inputs):

  case	myers	naive
  equal	3µs	3µs
  50 scattered edits	1.0ms	2.3ms
  100-line block moved	1.0ms	2.0ms
  disjoint (2k lines)	0.7ms	15.7s

• The old engine is renamed to what it is: naive (DiffNaive, greedy closest-match resync, not minimal). The flag/attr value unified remains a deprecated alias that now selects myers — matching its historic "like diff -u" promise — and enforces --format=unified (CLI check + bzl fail()).

3. Feature/algorithm support matrix

• New diff_test_formats_test macro crosses output formats (each format checked against its own golden) with algorithms; golden coverage now spans {myers,naive} × {unified,context,normal} plus direct × all formats.
• AlgorithmFeatureMatrix unit test proves all comparison options (ignore_case, ignore_*_space, ignore_blank_lines, ignore_matching_lines, strip_comments, regex_replace_*) behave identically under naive, myers and direct.
• Documented corner case (pinned by test): ignore_case + a case-sensitive ignore_matching_lines expression diverges between naive and myers; write such expressions as (?i)....
• max_diff_chunk_length only applies to naive (documented; myers uses its internal cost cap).

Test plan

• bazel test //... — 91 tests pass (40 in //mbo/diff/..., incl. 15 new golden targets and the new unit tests).
• MyersRoundTrip property test: apply(lhs, diff) == rhs over a deterministic corpus incl. a disjoint pair that exercises the cost-cap fallback.
• Manual: byte-comparison against system diff/diff -c/diff -U0 and patch round-trips for all three formats × {naive, myers} on 11 input families; hang-regression bisect for the fallback path (fixed grid-edge clamping in the furthest-reaching recurrence).