feat(diff): A/B-labeled change format + root-input comparison

[christophergeyer]

Summary

Two changes to roar diff output, both verified against a real MNIST pipeline:

• A/B-labeled change format. Param, code, and env diffs render A: x B: y instead of x -> y. The arrow left the A/B mapping implicit (inferred from the header) and implied a before→after direction that doesn't apply — A and B are siblings being compared, not a timeline.
• Root-input comparison. roar diff now compares the raw artifacts each lineage started from, matching roar inputs semantics (a producer with no inputs — a download / roar get — is itself a source). When root inputs match, UNCHANGED gains an inputs (same artifacts) line. When they differ, an INPUTS section lists the divergence and supersedes the per-step DATA category so the same fact isn't reported in both places.

New collect_root_inputs helper + root_inputs_match / root_inputs_a / root_inputs_b on GraphDiffComputation and DiffResult; JSON output includes the new fields.

Test plan

• tests/unit/test_diff_engine.py + test_diff_renderer.py — 14 new tests (A/B format, collect_root_inputs incl. the inputless-producer-is-root case, root-input comparison, renderer INPUTS/UNCHANGED/DATA-suppression, JSON fields).
• Full unit suite: 882 passed, 1 skipped.
• Manual: exercised all five diff dimensions (params, code, inputs-differ, pipeline-structure, identical) on a real MNIST extract → train pipeline.

Out of scope (pre-existing, observed during testing)

• The --output filename shows up as a PARAMS diff and can score as ROOT CAUSE over a real code change (PARAM_CHANGED 0.9 > CODE_CHANGED 0.85).
• A wget step whose URL differs past the truncation point renders "command differs" with identical-looking A/B.

Both are scoring/truncation papercuts, separate from this change.

🤖 Generated with Claude Code