fix: batch citations resolve against the universal ledger

[KadenMc]

Summary

In 0.6.0 the job-id citation check was correctly taught to consult both the local signac project AND .aexp/ledger/<id>.json, but the batch-citation check still went through aexp.linking.list_batches(), which walks the local signac project only. The result: a finding citing {type: batch, experiment_id: E005, selector: {condition: X}} spuriously emitted finding.empty_batch whenever the matching terminal runs lived only in the ledger (e.g., backfilled from another machine).

The fix is purely additive: build an (experiment_id, condition) -> [job_ids] lookup from ledger entries and count those matches in the batch-resolution branch before falling through to the elsewhere/empty paths.

Surfaced by

The electricrag laptop validate during the 0.6 migration. After the cluster backfilled 20 ledger entries (including the 3 E005 batch members F005 cites), the laptop validator still emitted 3 finding.empty_batch errors — exactly because none of the cited E005 jobs had local workspaces; they existed only in the just-pulled ledger.

No behavioral change for existing cases

• Local-only batch matches still resolve via list_batches.
• Elsewhere (per-machine index) matches still emit finding.absent_batch_runs warnings.
• Empty/broken batches with no matches anywhere still emit finding.empty_batch errors (or warnings under --strict-runs=warn).

Only new: ledger-only batch matches now correctly count as "here."

Regression test

tests/test_validate_cross_machine.py::test_validator_batch_citation_resolves_via_ledger_only writes a ledger entry with no corresponding local workspace and asserts a batch citation pointing at that entry's (experiment_id, condition) resolves cleanly.

The test's registered_machine field is documented inline as informational only — the validator does not consult it. This makes the Phase 2 design choice explicit: the ledger is universal; machine identity is provenance, not a query filter. (If we ever want machine-aware filtering for batch citations, it would be a selector extension — {type: batch, ..., selector: {condition: X, registered_machine: cluster}} — handled in _selector_matches. Separate feature, separate discussion.)

Version

0.6.0 → 0.6.1 (patch). PyPI history goes 0.5.0 → 0.6.1 since 0.6.0 was tagged in source but never published.

Test plan

• poetry run pytest tests/test_validate_cross_machine.py tests/test_validate.py tests/test_ledger.py — 69 passed locally
• poetry run ruff check src/ tests/ — clean
• After merge: on the consumer (electricrag laptop), aexp validate drops from 4 errors → 1 error (just the remaining d84ce97... never-ran job citation, which is a data issue, not a validator bug).