FIX Batch v1.0 conversation migration backfills

[romanlutz]

Description

The v1.0 upgrade path performs one database round trip per conversation while backfilling conversation and target-link data. At the production data shape, this leaves tens of thousands of SQL Server operations in both b2f4c6a8d1e3 and e5f7a9c1b3d2; the original migration has already rolled back after communication-link failures, and disposable SQL Server profiling showed the later target-link backfill becoming the dominant bottleneck after fixing the first one.

This change:

• inserts Conversations in bounded 400-row, 800-parameter statements;
• groups identical retained target identifiers so each graph is materialized once;
• stages conversation target links in the same bounded batches, then performs one SQL Server join update or one SQLite correlated update;
• preserves first-non-null target selection, conflict warnings, attack-only placeholders, existing-row idempotency, nullable pyrit_version, best-effort identifier handling, and the outer Alembic transaction;
• adds focused failure, rollback, fallback, cleanup, batching-limit, and idempotency coverage plus an opt-in production-scale SQLite upgrade test.

This intentionally modifies two committed migration revisions. Production is still before both revisions, so a new later revision cannot reduce the round trips that the upgrade must execute first. The Alembic immutability guard therefore requires an explicit release-maintainer exception for this hotfix.

Disposable SQL Server 2022 validation reduced the complete 9c8b7a6d5e4f to v1 head upgrade to 33.141 seconds. e5f7a9c1b3d2 fell from 398.68 seconds to 9.679 seconds. No production database was accessed.

Tests and Documentation

• 8 focused migration batching and fallback tests passed.
• Complete tests/unit/memory/test_migration.py: 47 passed.
• Production-scale SQLite upgrade: 135,069 prompt rows, 81,432 conversations, 204 conversation inserts, 204 target-link staging inserts, 1,227 total statements, 6.785-second migration runtime.
• Disposable SQL Server 2022 full upgrade and forced rollback validation passed, including staging-table cleanup.
• Changed-file pre-commit checks and git diff --check passed.
• Documentation: N/A; migration behavior is unchanged apart from bounded execution.