Separate the comparison engine from the reporter

[AnshBajpai05]

Second slice from #80 — the one you said you were most excited about. Pure refactor: no new flags, no behavior change, identical output.

What changes

compare_traces() used to compute and print at the same time, which meant any new output format would have to re-implement the comparison logic (or re-parse printed text). Now:

• compare.py is pure — no print, no side effects. It returns a richer ComparisonData:

  ComparisonData(
      deltas,             # every function, with a status
      regressions,        # unchanged: List[RegressionData]
      improvements,       # functions that got faster past the threshold
      new_functions,      # names only
      removed_functions,
      threshold,          # what it was compared against
      has_regression,     # unchanged
  )

  Each FunctionDelta carries name, status, old_time, new_time, percent. Statuses are module-level constants: regression / improvement / stable / new / removed / no_signal / no_baseline — the states the old code expressed only as printed strings.

• display.py is new and owns all rich printing: print_comparison(comparison, only_regressions=False). The show_only_regressions argument moved from the comparison function to the display function, which is where filtering actually belongs.

• cli.py now does compare_traces(...) → print_comparison(...). Both call sites (--compare and baseline compare) updated.

Why this ordering matters

This is the enabler for the Markdown reporter (next PR): with comparison results as data, a reporter is ~100 lines consuming ComparisonData, and the same is true for anything after it — JUnit, the HTML comparison column, whatever comes later. None of them need to know how a regression is decided.

RegressionData and has_regression keep their shape, so any external consumer of the current API keeps working.

Verification

• Existing behavior tests were split by concern: test_compare.py now asserts on returned data (including a test that comparison produces no output at all), and the printing assertions moved to a new test_display.py — same expectations, same strings, just pointed at the display layer.
• Added a format-fidelity test that pins the header, the arrow line, and the (+N.NN%) formatting.
• Manually diffed real --compare output before and after the refactor: identical, down to blank lines and colors.
• Full suite green (113 passing locally; the one failure on my machine is the pre-existing test_symlinked_user_code, which needs Windows symlink privileges and passes on CI).

🤖 Generated with Claude Code

[KaykCaputo]

Perfect, just as i imagined