refactor(validators): compose common validators centrally (supersedes #316)

[shujaatTracebloc]

Summary

The 0-record guard, TableNameValidator, DuplicateValidator, and LabelDiversityValidator were copy-pasted across the per-category factories — TableName+Duplicate in all 11, the 0-record guard wired three different ways (NLP _nlp_content_validators, vision _zero_record_validator, inline for tabular), label-diversity in each of the 6 classification factories. That repetition is exactly what let the tabular 0-record gap happen — a factory just forgot to add it.

This centralizes the universally-applicable validators in map_validators, driven by declarative ModalitySpec traits, so each factory returns only its category-specific validators.

Composition (one place)

[IngestableRecordsValidator(file_subdir=spec.file_subdir)]   # every category
  + spec.build_validators(options)                           # category-specific
  + [LabelDiversityValidator] if spec.is_classification
  + [TableNameValidator, DuplicateValidator]                 # every category

Changes

• New ModalitySpec traits: file_subdir (images/texts/sequences/None) and is_classification.
• map_validators builds the common frame around the factory output.
• Factories drop the repeated lines; deleted _zero_record_validator; _nlp_content_validators → _text_content_validator (text-content only); label_diversity_validator made public.

Behavior

Set-preserving — the resulting validator set per category is unchanged, except tabular_classification / tabular_regression / time_series_forecasting / time_to_event_prediction now correctly carry the 0-record guard. This closes the tabular gap centrally and supersedes #316. Only order changes: the 0-record guard now leads (fail-fast); pass/fail is order-independent.

Tests

Updated the order-sensitive mapping assertions; added test_common_validator_frame_composed_for_every_category that locks the frame (guard first, table/duplicate last, label-diversity for classification only). Full suite: 1226 passed, 1 xfailed, coverage 96.6% (gate 95%).

🤖 Generated with Claude Code

---

Note

Medium Risk
Touches the preflight validator chain for all 11 task categories and intentionally adds 0-record checks to tabular/time paths; regression risk is mitigated by broad mapping tests but empty-manifest failures will surface earlier for those categories.

Overview
Shared preflight validators (IngestableRecordsValidator, optional LabelDiversityValidator, TableNameValidator, DuplicateValidator) are no longer duplicated in each per-category factory. map_validators now wraps every ModalitySpec.build_validators output in a single frame, driven by new spec fields file_subdir and is_classification on ModalitySpec / registry.py.

Per-category factories in modalities/validators.py return only category-specific checks; _zero_record_validator is removed and NLP wiring is split so _text_content_validator handles UTF-8 content while the 0-record guard is built centrally from spec.file_subdir.

Behavior: Validator sets stay the same for most categories, but tabular / time-series families now get the 0-record guard (header-only / empty CSV), closing the gap called out in #316. Order changes so the guard runs first for fail-fast; pass/fail is otherwise order-independent.

Tests update expected ordering and add test_common_validator_frame_composed_for_every_category to lock guard-first, table/duplicate-last, and label-diversity only on classification categories.

^{Reviewed by Cursor Bugbot for commit e20467a. Bugbot is set up for automated code reviews on this repo. Configure here.}