[alex-mextner]

Refs #890

Main fails the universal Rig lint after the real-history KPI gate landed (#827, merged as 366918f): 31 ruff-check errors, one ruff-format file, and two handwritten files over the 1200-line warning limit. This PR is one atomic repair of all three, so main needs one release instead of two.

What changed

• Ruff repair of the four real-history Python files: the same commit as the lint-only PR fix(lint): repair ruff findings in scripts/real-history (Refs #891) #892, cherry-picked unchanged (follow-up ticket Repair ruff findings in scripts/real-history from the intent engine PR #891). It is not duplicated work: if that PR lands first, this one rebases onto it.
• scripts/real-history/build_real_corpus.py (1461 lines) is split into corpus_sources.py (inputs and rows), corpus_turns.py (entities, profile, turns), corpus_gates.py (output gates) and the CLI entry build_real_corpus.py. The entry name and arguments are unchanged.
• src/services/intent/seed-calendar.ts (1347 lines) is split into seed-calendar-vocab.ts, seed-calendar-read.ts, seed-calendar-free.ts and the assembling seed-calendar.ts. The only importer, seed-catalog.ts, is unchanged.
• One small defect fix found in review: OpenedDb now removes the unpacked temp copy when a corrupt .gz or a failed open stops it half-way (private chat data stayed in the system temp directory). A synthetic regression test covers it.

Behaviour proof (no corpus, no database, no benchmark touched)

• Seed: the serialized calendarFamilies (12 families, names, field order, pattern hashes) has the same sha256 before and after: 20471407d4da23db97213ca7d8f53e5798beb719475d7b141cdfc0ee8021c836. No seed text, ID, link, version, model or KPI changed; no benchmark result file touched.
• Python: the AST of all 109 top-level definitions is identical before and after, except _scan_record, which is now scan_record because gates and entry live in different modules (no other caller exists).
• Synthetic end-to-end run of the split pipeline (scan, audit, build, verify on invented SQLite, gzip and log fixtures in a temp dir) gives the same canonical hash as the lint-only tree: 2a69f0f7158c2526220f3748e5f6edbae940d6ae50012d7dc88c84f058215b41. Input bytes unchanged, no residual temp files, same exit codes.

No lint suppression, config change or threshold change. uv.lock (1343 lines) stays a known warning.