test: specify extract_batch contract

[laynepenney]

Summary

• Specify the Python extract_batch contract before implementation.
• Add Atlas's SHA-pinned real Ministral-3B failure fixture pack from config#480.
• Lock structured BatchUnit input, injected inference, per-unit attribution, strict shaping and validation, fallback, and the output count invariant.

Contract covered

• BatchUnit(id, text, capabilities?) input and source_unit_id output attribution
• provider-neutral injected infer seam with no recall or MLX dependency
• exact output cardinality and ordered per-unit terminal slots
• failure reasons unparseable, schema_invalid, dropped, and merged
• outer Markdown fence and JSON-aware // comment stripping
• narrow shape coercion and unknown-key removal
• capability-scoped entity_refs: coerce with entities in scope, drop for recall's set
• explicit temporal prompt/schema normalization to raw and resolved
• standard call default plus per-unit capability overrides

Real fixture provenance

Copied verbatim from synapt-dev/config config#480, merged at 43c4687:

design/results/extract-batch-limits-2026-07-13/fixtures/extract-batch-real-failures-v1.json

SHA-256: 9b183f18ab5116cfb1f5ee67d0e99cd5af3fb7f7b99d649b1d58821f9e7489f1

The pack covers 21 raw responses, 25 malformed leaves, 10 dropped source occurrences, 2 unknown-key leaves, 3 temporal conflict cases, and 1 explicitly contract-derived scalar entity_refs case.

Verification

• python3 -m compileall -q tests/python/test_extract_batch.py: pass
• python3 -m pytest -q tests/python --ignore=tests/python/test_extract_batch.py: 307 passed
• python3 -m pytest -q tests/python/test_extract_batch.py: expected TDD red at _strip_output_hygiene() NotImplementedError in the landed batch skeleton (1 passed, then implementation gap)
• git diff --check: pass

Boundary: OSS extraction reliability only. No identity, org, evaluation, training, repair-product, or premium implementation logic.

Reviewers: Opus for the pinned contract seam; Apollo for the implementation contract-read.