[gabrielspadon]

Change

The scheduled monthly refresh now processes only Gabriel Spadon (Scholar bfdGsGUAAAAJ, DBLP 192/1659) instead of all 64 enabled authors in data/input.csv.

• data/website-authors.csv: the input.csv header plus the Gabriel Spadon row, copied verbatim.
• .github/workflows/monthly-refresh.yml: a workflow-level MONTHLY_CENSUS points at that file. python3 main.py takes --input "$MONTHLY_CENSUS" (line 137), and the manifest step derives enabled_authors and authors_with_bib from the same file (lines 205-217), so both counts describe one census.
• data/input.csv, the CLI default, local runs and refresh-shadow.yml are unchanged.
• The finalization tail and run_all are scoped to the authors in the given input (below), because --input against the standing output/ tree was not safe.

Data-safety analysis

What a one-author --input did to other authors' directories on main (7ee0ca5):

Pass	main	Effect on an out-of-scope author
orphan sweep	io_utils.py:379, postrun.py:175	scanned every directory, deleted duplicate orphans
year window	postrun.py:219	deleted any .bib below get_min_year()
post-run fixup	postrun.py:269	rewrote and renamed any .bib not in serializer form
superseded preprints	postrun.py:460	deleted preprints with a published twin
DOI conflict gate	io_utils.py:87, postrun.py:363	raised FinalizationError on a conflict in a directory the run cannot re-enrich
forced re-enrichment	scheduler.py:352, :361	counted out-of-scope copies, forcing passes that can never settle
a2i2 rebuild	io_utils.py:~663	emptied and rewrote output/a2i2/ when at least half the members resolved

Not affected, read and confirmed per author directory: pipeline/article.py and merge_utils.py removals (author_dir is derived from the record being processed), the canonicalize.py post-run orphan repair (a per-entry rule inside _fixup_bib_entry, no file removal), and reconcile_summary_csv (drops only rows whose file is already gone).

After this change every pass above reads only the directories of the given records (iter_output_dirs(only=...), fsscan.py:276). baseline.json still mirrors the whole tree and is a derived file. build_a2i2_folder rebuilds a populated folder only when every member resolved (io_utils.py:661), replacing the half threshold.

Evidence, both from a copy of the committed output/ (3,693 files), one-author finalize_run:

• main code: FinalizationError: conflicting copies of the same DOI remain in author directories: Haque (PMTOOSQAAAAJ), Oore (cI0dYX4AAAAJ), Trappenberg (EwkaTYEAAAAJ). The committed corpus already carries these conflicts, which only a full run's forced re-enrichment closes, so the monthly job would have failed finalization every run.
• this branch: finalization completes, a2i2 rebuild skipped: only 1 of 21 members resolved ... which would delete 1105 existing file(s), and sha256 over every file under output/ is identical before and after (0 differences).
• hazard tree (tests/test_finalize_run.py), main code with the conflict gate disabled so it reaches every pass: 3 out-of-scope files deleted (Roe2019-Old.bib, Roe2022-Detecting.bib, Roe2024-AlphaCopy.bib), 4 rewritten. Both new finalize tests fail on main and pass here.

Consequence to accept: a cross-author DOI conflict between Gabriel's copy and a co-author outside the census is no longer visible to a one-author run. The full run still gates it.

Website consumer evidence

MAPS-Lab/mapslab-website .github/workflows/import-publications.yml (blob d93ac20, main 1a26cbf) checks out MAPS-Lab/CiteForge with sparse-checkout: output/Spadon (bfdGsGUAAAAJ) and runs scripts/aggregate_bib.py "_citeforge/output/Spadon (bfdGsGUAAAAJ)/". No other directory is read. test_the_website_census_is_the_one_author_the_website_reads pins the census to that directory name.

Test record

• Command: heavy .venv/bin/pytest -m 'not live' --cov=citeforge --cov=main --cov-fail-under=68 -q, the tests.yml selection plus the 3.12 coverage leg.
• Commit 2687e52, host RTX-4090, Python 3.14.4.
• 2091 passed, 0 failed, 2 skipped, 7 deselected (live). Coverage 84.15 percent against the 68 floor.
• The 2 skips are the pre-existing cutover contracts at tests/test_workflow_contracts.py:308 and :339.
• ruff check citeforge/ tests/ main.py clean, mypy citeforge/ main.py clean (56 files).
• Not run locally: the 3.10 and 3.12 interpreters (both byte-compile the touched files), a live monthly run (needs provider keys), and uv pip compile lock freshness (no dependency changed).

New tests: test_finalize_run.py:923 (byte-identical tree, summary CSV unchanged, in-scope sweep still runs), :967 (a2i2 majority guard), test_workflow_contracts.py:203/224/230 (every main.py call and the manifest read the census; census is the website directory and a verbatim input.csv row), test_scheduler_parallel.py:238, test_regression.py:2432, test_fsscan.py:44. Existing callers updated for the required author_dirs argument, and the existing unresolved-conflict test now passes both authors so it still exercises a full-scope abort.

Left for the owner

• output/a2i2/ is no longer refreshed by the monthly run, since 1 of its 21 members is in scope. It stays at its last full-run content until citeforge runs on the full census.
• The other 63 authors' committed output stops moving monthly.