[derek73]

Follows #854, which deferred this split. tools/differential/local_corpora.py (2,374 lines) becomes:

• local_core.py: what the sources share (Fetched, Source, the retry policy, the share gate, _download, fetch_command, and an unchecked directory)
• one module per source: local_ctgov.py, local_fec.py, local_tse.py, local_tw.py, local_ukec.py, local_psc.py, each opening with the description of its source that used to be in local_corpora's docstring
• local_corpora.py (about 820 lines): the SOURCES registry, the cache's files (fetch, load, read_records), the views, the word index, the fingerprints and the command line

Every documented tools/differential/local_corpora.py … command and every load_tool("local_corpora") works unchanged, and so do the imports in name_usage.py, snapshot.py and local_agreement.py. No behavior moves.

Commits, for blame (merge without squashing)

1. The move only. Line ranges are cut verbatim into the new files; the only new lines are each module's docstring head and imports. git blame -C -C on a moved line reaches the commit that wrote it (for example _backed_off → Simplify the local corpora: one retry policy, one share gate, shared record and fingerprint helpers #854, the fec-individuals paragraph → Add the FEC's individual contributors as a local corpus #834). This commit does not pass the suite on its own, because the tests still patch the old module. Its message says so.
2. The edits. Repointing the test patches, qualified reads in local_tse, a guard test, and docs.
3. From the review. Each constant is bound in exactly one module, the guard checks every module pair, and moved comments that named something now in another module say which.

A squash merge would fold the move into the edits and lose the blame trail, so this needs a merge commit or a rebase merge. If a later git bisect lands on d666bb9, it needs git bisect skip.

Monkeypatches

A patch only reaches a fetch when it is set on the module whose globals the fetch reads. So the tests' 43 patch sites on a moved constant now target that module: local_core for RETRIES and CHUNK, local_fec for FEC_MAX_BAD, and so on. The test module takes these modules from sys.modules, because a second load by path would be a copy that no fetch reads. local_tse reads local_core.RETRIES, USER_AGENT and _TRANSIENT through the module, as _backed_off does, so one patch reaches every reader.

No module from-imports an upper-case name from a sibling module, except PSC into local_corpora. local_corpora's credit lines read FEC_CYCLE and TSE_ELECTIONS through their modules. As a result, a stale setattr(local_corpora, "TSE_MAX_BAD", …) raises instead of passing with its guard off. test_no_module_imports_a_sibling_constant holds this over the modules SOURCES names, so a new source is covered. Its negative controls fail it: importing USER_AGENT into local_tse again, or FEC_CYCLE into local_corpora.

Verification

• Negative controls retaken. Each of the 43 repointed patch sites was removed one at a time and the test module run, on master (48ed448) and on this branch. All 43 sites give the same failing tests on both trees: 35 fail tests when removed, and 8 made no test fail before the split either (those are pre-existing, not touched here).
• Fingerprints unchanged. fold_fingerprint() and local_agreement.scorer_fingerprint() are byte-identical to master, so cached word indexes and the PSC agreement record stay valid.
• Same answers on all six cached corpora (hashes only, no names printed): records, meta, header(), tally, cached and freshly built word index, lookups, pair_counts for tse-donors, and name_usage.py's local rows are identical to master. So is the output of local_corpora.py fetch <source> (cached path) and pairs tse-donors.
• Read back through the fetch functions. fetch_fec() and fetch_psc(), run on their cached inputs with the network refused, give records whose sha256, detail and data date match meta.json. tw, ukec, tse and ctgov fetch from the network, so they were checked only through the comparisons above.
• local_agreement.py psc-individuals matches its record.
• ruff, mypy and the full suite pass on 3.11 (23,095 passed). The local-corpus, agreement, name_usage and snapshot test modules pass on 3.15.

One pre-existing finding, not from this branch: snapshot.py --local ukec-individuals reports 508 moves on master as well, with identical output. The cached local snapshot predates recent title changes and needs re-recording.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.05%. Comparing base (48ed448) to head (1fe3fad).

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #855   +/-   ##
=======================================
  Coverage   99.05%   99.05%           
=======================================
  Files          47       47           
  Lines        4669     4669           
=======================================
  Hits         4625     4625           
  Misses         44       44           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.