cowork-bot: automated improvements (cowork/improve-json2sql-20260808)

[github-actions]

Automated improvement PR from the Cowork repo-improver rotation (one coherent senior-dev improvement per run; see individual commit messages). Subsequent runs push additional commits to this PR rather than opening new ones.

[Coding-Dev-Tools]

Pre-PR Code Review: REQUEST_CHANGES

Reviewer: Pre-PR Code Analyzer
Date: 2026-08-08
Verdict: REQUEST_CHANGES

Summary

This PR improves generate_schema to infer primitive column types (INTEGER, DOUBLE, etc.) instead of hardcoding TEXT for primitive arrays. The change is well-tested with 3 new regression tests covering int/float/bool arrays across PostgreSQL, MySQL, and SQLite dialects.

Positive Findings

• Type inference for primitive arrays is a genuine improvement
• Tests cover all three SQL dialects with appropriate expected types
• The formatting cleanup (removing unnecessary line wraps) improves readability
• CI passes (ensure-pr workflow green)

Concerns

1. Edge case: mixed-type primitive arrays
The inference uses sql_type_for(data[0], ...) — it only looks at the first element. If the array is [1, "hello", 3], the schema would declare INTEGER but the second value would fail to insert. Consider:

• Sampling more elements (first N or all)
• Falling back to TEXT if types are heterogeneous
• Documenting this limitation

2. Empty array handling
col_type = sql_type_for(data[0], ...) if isinstance(data, list) and data else "TEXT" — the else "TEXT" covers empty arrays correctly, but the conditional is dense. A comment would help.

3. Age/contributor gate
This PR is less than 6 hours old and has only 1 commit from 1 contributor. Per the review policy, PRs need either 6+ hours of age or 3+ distinct contributors before approval. This will be re-evaluated on the next rotation.

Required Changes

1. Consider adding a test for mixed-type primitive arrays (or document the single-element inference behavior)
2. Wait for the 6-hour age gate or additional review passes before merge