feat: add sentence_pair_classification modality

[shujaatTracebloc]

What

Enable the NLP sentence-pair classification use case end-to-end on the ingestor side — Phase-4 cross-repo wiring for engine tracebloc-engine#301.

The on-disk contract: one .txt per sample under texts/, each a tab-separated text_a<TAB>text_b sentence pair, plus a class label in the labels CSV.

sentence_pair_classification is supervised text classification — the label travels in the CSV exactly like text_classification (is_classification=True, is_self_supervised=False, staged from texts/) — but the .txt has a structured shape (the text_classification layout with a tab separating the pair).

Wiring (mirrors text_classification; structural check mirrors embeddings)

• utils/constants.py — TaskCategory.SENTENCE_PAIR_CLASSIFICATION.
• schema/ingest.v1.json — enum value, texts requirement + description, and added to the "requires label" rule (supervised — deliberately not the self-supervised no-label guard).
• modalities/registry.py — ModalitySpec (file-bearing, classification, NLP, texts/). is_nlp gates the #805 data-derived text profile (the ingestor side of the federated tokenizer-alignment check — there is no check_same_tokenizer in this repo; that lives in backend/SDK).
• modalities/validators.py + transfer.py + cli/conventions.py — same validator set + raw-text staging from texts/ as text_classification.

What's distinct from text_classification

The on-disk shape is structured, so beyond the shared TextContentValidator (UTF-8/binary hygiene) it adds a structural check that rejects any .txt that isn't exactly 2 non-empty tab fields (no plain prose, no empty side, one record per file).

Rather than clone ContrastivePairsValidator, this extracts the shared scaffold into a new TabSeparatedRecordValidator base (centralized validators, #317). ContrastivePairsValidator (2-or-3 fields) subclasses it with byte-identical messages — its existing tests pass unchanged — and SentencePairValidator (exactly-2) is a thin subclass.

Template fails loudly via the shared run_ingestion helper (no exit 0 on hard failure); zip(ids, batch) send path untouched.

Tests

• New templates/sentence_pair_classification/ (script + README + labels CSV + 5 sample pairs), examples/yaml/sentence_pair_classification.yaml.
• New tests/test_sentence_pair_validator.py + tests/test_sentence_pair_classification.py (supervised — incl. rejecting a config that omits label).
• Updated the enumerating/contract tests (registry, validators_mapping, schema_validation, template_equivalence, modality_failure_accounting) + e2e params.
• Unit suite green locally (1353 passed, 1 xfailed); black-clean. e2e runs on CI (MySQL service).

🤖 Generated with Claude Code

---

Note

Medium Risk
New modality touches schema, registry, validators, and ingestion paths used by all YAML submissions; behavior is well-tested but expands the supported category surface customers rely on.

Overview
Adds supervised sentence_pair_classification end-to-end: CSV manifest + texts/ sidecars where each .txt is exactly text_a<TAB>text_b, with class labels in the CSV like text_classification.

Wires the new category through TaskCategory, ingest.v1.json (enum, required texts + label), modality registry (NLP, file-bearing, classification), conventions, transfer (text_transfer from texts/), and a validator stack that adds SentencePairValidator on top of the usual text hygiene and label checks.

Introduces shared TabSeparatedRecordValidator and refactors ContrastivePairsValidator to subclass it (embeddings pair/triplet vs sentence-pair exactly-2 fields). Ships template, example YAML, docs, e2e/characterization cases, and broad test coverage; bumps package version to 0.5.5.

^{Reviewed by Cursor Bugbot for commit dc356b1. Bugbot is set up for automated code reviews on this repo. Configure here.}