feat(metrics): add Recovery Step Ratio (RSR) for post-error recovery effort - #6

What

Adds the Recovery Step Ratio (RSR) metric (roadmap B2): how many successful steps a run spends recovering after ERROR/RETRY/ABANDONED clusters, candidate vs baseline. DiffReport gains baseline_recovery_steps, candidate_recovery_steps, recovery_step_ratio; the gate is opt-in via --max-recovery-ratio / max_recovery_step_ratio.

Why

Roadmap item B2 — context/ROADMAP.md WS-B: "Recovery Step Ratio (steps to recover from error vs baseline)". WEI says a run failed; RSR says how expensive getting back on track was — a run can fail cheaply and recover fast, or burn many steps limping back to the golden path.

How

• Semantics: a recovery window opens at each wasted-status step; successful steps inside it count; the window closes at the first step re-aligning with the other trace (MATCHED/MODIFIED — TDI's LCS set). Unrecovered tails count fully. Failed steps themselves stay WEI's territory.
• Ratio: cand/base when baseline needed recovery; raw candidate count when the baseline is clean (nothing to be relative to); 0.0 when both clean.
• Surfacing: summary(), terminal panel, PR markdown (gate row only when threshold provided → default output unchanged), and a new --explain finding category recovery.
• Gates: opt-in only (None default) in assert_no_regressions, CLI flag, [cli] max_recovery_ratio, [assertions] max_recovery_step_ratio — zero behavior change for existing users.
• Additive defaulted report fields → JSON contract stays backward compatible.

Testing

• 22 new tests in tests/test_recovery_steps.py: metric edge cases (no errors, immediate retry, added-steps-then-rematch, trailing unrecovered, multi-cluster, both sides), ratio cases, comparator integration, JSON round-trip with new fields, summary block, opt-in gate behavior, explanations, config plumbing.
• Full suite: 227 passed (was 205); make lint + format clean.
• Manual e2e: CLI run shows Recovery Steps (base/cand): 0 / 2 (RSR 2.00), explanation finding, gate FAILED at --max-recovery-ratio 1.0.

Checklist

• Exactly one logical change in this PR
• make lint passes
• uv run python -m pytest is green (227 passed)
• Docs updated if user-facing (README metrics table)
• CHANGELOG.md entry added under [Unreleased]
• Conventional title