[PXT-477] Add migration for existing primary key tables

[Chrisclone]

Migration Strategy for Existing Primary Key Tables

Implements the migration path for tables created with primary_key columns before the PK enforcement PR.

Problem

PR pixeltable#1203 adds a partial unique B-tree index for new tables, but existing tables with is_pk=True columns have:

• No PrimaryIndexMd in their metadata
• No physical unique index in PostgreSQL
• Potentially duplicate live rows (the bug this PR fixes)

Solution: Schema Version 48 → 49

Converter convert_48.py runs during catalog load and for each base table with PK columns but no index:

1. Generates PrimaryIndexMd (matching create_initial_md() logic)
2. Attempts CREATE UNIQUE INDEX inside a savepoint
3. If succeeds → persists updated metadata with primary_index_md, PK enforcement is active
4. If fails (duplicates exist) → rolls back savepoint, strips is_pk from all columns, logs warning:

   "Table X had primary_key columns with duplicate rows. Primary key constraint removed."

Design Decisions

• No silent deduplication — removing PK designation is safe; user data is untouched
• Savepoint isolation — failed index creation doesn't abort the migration transaction
• String PKs use left(col, 256) truncation matching BtreeIndex.MAX_STRING_LEN
• Views skipped — only base tables get the physical index
• Idempotent — checks for primary_index_md is None before running

Files

• pixeltable/metadata/converters/convert_48.py — migration logic
• pixeltable/metadata/__init__.py — version bump 48 → 49
• pixeltable/metadata/notes.py — version note
• tests/test_pk_migration.py — 4 test cases (happy path, duplicates, composite PK, idempotency)

---

Generated by OpenClaw