feat(mcp_readability): decide what a run can reuse from its baseline

[akangsha7]

Stacked on #637, which moved these modules into scorers/mcp_readability/. Follows #622 and #615, both now merged.

Summary
#622 records the judge's inputs and #615 reads the previous run's back. This is the logic between them: given a baseline and this run's fingerprints, decide which tools still hold, which must be re-judged, and what explains the difference. Pure functions — nothing calls them yet, so a run still judges every endpoint in full.

The invariant: same endpoint, same per-tool fingerprints, same judge fingerprint means byte-identical feedback and zero model calls. Every departure gets a named reason from an exact set-difference, not a guess.

Changes

1. scorers/mcp_readability/carry_forward.py (new) — decide returns full, partial or carried for one endpoint; decide_with_components names which judge input changed. A changed judge fingerprint invalidates every finding, not just the ones for changed tools.
2. Merging — merge_feedback takes the judge's entries only for changed or added tools and carries the rest. The judge is still shown the whole man page, since its severity calibration is relative to the full surface, so the filter rather than the prompt is what provides the guarantee.
3. Finding identity — mint_finding_ids gives each finding a stable id that survives carrying, which is what makes new versus resolved countable. Tool plus rule_id is not unique, so a locator and the normalised title disambiguate.
4. Provenance — build_provenance records where each tool's findings came from, which are new versus resolved, and the baseline job and timestamp being compared against. Persisted to mcp_readability_feedback_provenance_json only.
5. test/mcp_carry_forward_test.py (new) — covers both failure directions: carrying a finding that should have been re-judged, and re-judging a tool nobody touched.

[helloeve]

It seems to me that MCP readability scorers are having multiple modules now, can we move the dependent modules into a dedicated directory under scorers? (similar to the dataset_quality one) and under scorers there is a top-level mcp-readability scorer which reference those modules?