Fix silent data loss in the Prometheus pipeline: raise on missing tables, sync TRUTH.PROMETHEUS

[sevmag]

Closes #909.

Both failure modes in the Prometheus conversion/training path were silent — truth-less or photon-less datasets with no error. This PR makes the first one loud and fixes the second at its root.

Changes

PrometheusReader: raise on missing tables. The reader dispatched extractors by exact table-name match and silently skipped any extractor whose table wasn't in the file (e.g. a non-default photon_field_name in the Prometheus config, or a typo'd table_name). It now validates all configured tables once per file and raises a ValueError listing the missing and the available tables.

TRUTH.PROMETHEUS synced with current Prometheus output. The constant still described a truth schema (injection_*, primary_lepton_1_*, ...) that Prometheus stopped writing long ago; against a freshly converted file, Dataset._remove_missing_columns silently stripped every truth column. Now:

• TRUTH.PROMETHEUS = the event-level truth written by every current Prometheus injection type (interaction + initial_state_*), matching Injection.to_dict().
• The old list is preserved as TRUTH.PROMETHEUS_LEGACY for older files such as the bundled example database (prometheus-events.db).
• PrometheusTruthExtractor derives its default columns from TRUTH.PROMETHEUS, so the constant and the extractor can no longer drift apart.
• PrometheusTruthExtractor gained an optional columns override for truth fields that only some injection types write (e.g. LeptonInjector's bjorken_x, bjorken_y, column_depth — deliberately not in the default, since GENIE-based files don't have them).
• All in-repo consumers point at the schema matching their data: the six examples/04_training scripts and the datamodule tests (which run on the bundled legacy database) use PROMETHEUS_LEGACY; PublicPrometheusDataset._event_truth now derives from TRUTH.PROMETHEUS instead of duplicating it.

Breaking changes

• TRUTH.PROMETHEUS changes meaning: code using it against old databases must switch to TRUTH.PROMETHEUS_LEGACY. This is intentional — the default-looking name should describe what current Prometheus writes.
• Prometheus files containing no photon table now raise in PrometheusReader instead of quietly converting to truth-only databases. Note this includes files from runs where zero events produced hits (Prometheus writes mc_truth-only files in that case) — previously these converted silently; failing loudly surfaces e.g. vertices injected outside the detector volume.

Testing

• New tests/data/test_prometheus_reader.py: extraction of both tables from the shipped Prometheus test parquet, verification that every TRUTH.PROMETHEUS field exists in that (current-schema) file, the columns override, and the missing-table raise.
• tests/data/test_datamodule.py + the new tests: 14 passed locally; pre-commit (black, flake8, docformatter, pydocstyle, mypy) green on all touched files.

🤖 Generated with Claude Code