# Task 04 — Schema Profiling & Fingerprints

## Goal
Capture the record structure of every CSV without full load; compute schema fingerprints.

## Input
- `manifest/inventory.jsonl`, `manifest/tokens.jsonl`; sample sizes from `pipeline.json`.

## Actions
1. Write `scripts/profile_schema.py` (streaming; head N + tail N + random M rows only).
2. Per file capture: delimiter, header presence, column names (raw, ordered), column count, inferred dtype per column, null-rate in sample, min/max of numeric columns in sample, total row estimate (size/avg-row-len).
3. Fingerprint = hash of normalized ordered column-name list (lowercased, trimmed). Also compute a name-insensitive positional fingerprint (dtype sequence) for headerless files.
4. Emit `manifest/schema_profiles.jsonl`; add table `schema_profiles` to `manifest.sqlite`.
5. Log unreadable/malformed files to `manifest/profile_errors.csv`.

## Intermediate Result (done when)
- 100% of CSV files have a profile row or an entry in `profile_errors.csv`.
- Every profile carries a fingerprint.