fix(data_contract): enrich in bounded units so wide schemas and wide tables don't truncate

[mischa76]

Summary

Fixes the pre-existing data_contract truncation blocker flagged in the WP9 milestone: the enricher drafted all assets in one ForcedToolCaller call capped at max_tokens, so a wide schema's combined enrichment exceeded the cap and WP3's max_tokens guard raised LLMCallError — the full messy_insurance run could never complete end-to-end.

DataContractAgent.run now enriches in bounded units, scaling in both dimensions:

• one asset per call → a wide schema (many tables) never overflows;
• a table wider than _FIELDS_PER_CALL (40) is split by field into ceil(cols/40) calls → a wide table (many columns — routine in legacy insurance/banking sources, e.g. 256 attributes) never overflows either. _merge_enrichment folds an asset's chunks back together (merging fields, keeping the first doc).

The system prompt (carrying the full declared schema) is byte-identical across calls, so WP3 prompt caching makes the extra calls cheap on input tokens; the per-call output budget is raised to 8192 (output tokens are billed per generation, so headroom is free).

Verification

Live (real Anthropic API):

• 5-table / 38-column messy_insurance — 5/5 contracts, all 38 fields typed, 0 undetermined-type flags.
• A real 256-column table — 7 bounded calls (ceil(256/40)), all 256 fields contracted, no truncation.

Both cases previously raised LLMCallError.

Keyless (CI, no API key):

• test_enrichment_is_batched_one_asset_per_call — one bounded call per asset.
• test_wide_table_is_chunked_and_fully_enriched — 256 cols → chunked, every field covered.

331 tests green, ruff check . clean, uv run mypy clean (32 files).

Notes

• A 256-column table is now 7 sequential calls (correct and bounded). If very wide tables ever matter for latency, a future optimization could parallelize the chunks at some prompt-cache-warmup cost — not needed now.
• The data_contract agent was the specific blocker to a full messy_insurance run; this closes it. A full end-to-end vault-agent run on messy (all LLM agents, token-heavy) was not run here.

🤖 Generated with Claude Code