Mainframe corpus manifest with fetch/scan commands (pinned refs, cached DBs, a real BMS corpus) #3213

Part of #3122. Tooling item 4 from the #3120 retrospective.

Problem

The corpora named in #3120/#3122 are not on disk. The issue's training_repos/cobol_corpus/ doesn't exist in a checkout, and language-crucible has only 8-file excerpts. Getting them needed:

• cicsdev/cics-banking-sample-application-cbsa from branch July2024Refresh, because its main was emptied at the 2024 sunset;
• a scan with GITGALAXY_DISABLE_GIT_HISTORY=1, plus PYTHONPATH pointing at the worktree;
• copies to protect the originals from the refractor (cobol-refractor rewrites NEXT SENTENCE in the target repository's source files in place #3206).

Proposal

• A committed manifest of mainframe corpora: (name, url, ref, languages).
• fetch: shallow clone at the pinned ref into a cache directory.
• scan: build the master DB via galaxy_ir.scan_to_db, cached per engine commit.

Used by the answer-key scorer, the differential and the snapshot test.

Add a third corpus with real .bms maps. #3122 says the BMS Phase-2 row rests on no real-corpus evidence yet. CBSA's src/base/bms_src/ has 10 maps and may already cover it; the survey is part of this issue.