refraction differential: carddemo full-corpus baseline is stale (157 recorded, 15 measured) #3342

Problem

tests/cobol_mainframe/refraction_differential_baseline.json records 157 unexplained deltas for the full aws-mainframe-modernization-carddemo corpus. A local run on 2026-09-22 measured 15:

aws-mainframe-modernization-carddemo [full]: unexplained 15 (baseline 157) — improved (lower the baseline)

The gate fails only on increases, so a stale-high baseline hides up to 142 regressions on that corpus.

Fix

Fetch and scan the pinned corpora, run python tests/tools/refraction_differential.py --update-baseline --corpus aws-mainframe-modernization-carddemo, and attribute the drop in the PR to the merged PRs that explain it: the ones between the baseline's last bless and now.

This is local-only: CI doesn't clone the full corpora. Found while gating #3250 (PR #3306).