feat: schema introspection command and agent mode

[SpollaL]

Summary

• sentinel schema <file>: new subcommand that emits column names, types, null counts, unique counts, and min/max (for numeric columns) as JSON — lets agents understand the data before writing rules
• --agent flag (or SENTINEL_AGENT=1 env var): forces JSON Lines output with row_count and duration_ms on every rule result, plus a structured summary line at the end
• Structured JSON errors on stderr in agent mode ({"type":"error","code":"...","message":"..."})
• CLI refactored to subcommand structure: sentinel validate (existing behavior) and sentinel schema (new)

Agent workflow

```bash

1. Understand the data

sentinel schema data.csv

{"columns":[{"name":"id","type":"int64","nulls":0,"unique":1000,"min":1.0,"max":1000.0}],"row_count":1000}

2. Validate with machine-readable output

sentinel validate data.csv -r rules.yaml --agent

{"type":"result","rule":"not_null_id","status":"pass","violations":0,"total":1000,"row_count":1000,"duration_ms":45}

{"type":"result","rule":"email_format","status":"fail","violations":12,"total":1000,"row_count":1000,"duration_ms":67}

{"type":"summary","passed":1,"failed":1,"quality_score":0.5,"duration_ms":312}

```

Test plan

• test_introspect_basic — column names and types correct
• test_introspect_nulls — null count correct
• test_introspect_unique — unique count correct (including duplicates)
• test_introspect_numeric_min_max — min/max correct for int columns
• test_introspect_row_count — row count matches
• test_agent_result_serializes_correctly — result JSON has all required fields
• test_agent_summary_serializes_correctly — summary JSON correct
• test_agent_error_serializes_correctly — error JSON correct
• test_agent_result_is_valid_json_lines — each line independently parseable

🤖 Generated with Claude Code

[SpollaL]

Merged manually into master via rebase (d85f7d6).