Add SGD context extraction

[aryamantodkar]

Summary

• add a validated SGD train context iterator for USER turns
• retain exact character offsets, slot values, service provenance, and context-only turns
• validate dialogue structure against train/schema.json
• reject unknown services, categorical span annotations, malformed offsets, duplicate dialogue IDs, and non-train splits
• preserve valid overlapping SGD annotations for the later ontology-mapping stage
• add focused tests for extraction and failure modes

Why

SGD supplies conversational context, not ITN labels. Enrichment needs a deterministic extraction seam before any service/slot pair can be approved for a SpanKind or combined with trusted donor values.

Invariant protected

This module reads only SGD train data and yields USER turns in deterministic file/dialogue/turn order. Every emitted span must reference a schema-declared non-categorical slot, exactly slice its utterance, and retain its source provenance. It deliberately does not create BIO labels or infer ITN classes.

Without this invariant, malformed offsets, categorical values, dev/test leakage, or guessed slot semantics could enter enrichment and recreate the semantic-label problems found in Google PHONE data.

Full-train verification

• dialogues: 16,142
• USER turns: 164,982
• non-categorical slot spans: 66,603
• context-only USER turns: 114,951
• two valid nested annotation cases were preserved

Validation

• focused tests: 11 passed
• full Python suite: passed
• Ruff lint: passed
• targeted Ruff format check: passed
• package build: passed
• Rust tests: 21 passed

The repository-wide Ruff format check still reports two pre-existing formatting differences in tests/test_golden.py and tests/test_types.py. Neither file is changed by this PR.

Scope

This PR does not add the audit CLI, ontology policy, donor pools, example generation, BIO compilation, dataset combination, splitting, or training.