Read export data, fidelity and validation from one native snapshot - #21

Phase 3 — consistent, database-read-only export

Keeps existing public commands, signatures and result shapes. Uses native database snapshots, not dataset copies, application locks, new configuration or dependencies.

Changes

• One private read boundary starts before catalog verification and owns connection/transaction cleanup.
• SQLite uses ordinary BEGIN; PostgreSQL/MySQL/MariaDB use repeatable-read isolation; Dolt additionally starts its native transaction explicitly despite retained autocommit.
• SAV export reads values, dictionary and import fidelity diagnostics on the same snapshot, resolving fidelity by the dataset UUID.
• Validation reflects its physical columns on that same connection rather than another engine.
• Connections are closed and owned engines disposed before SAV writing/publication and on failures. Removing the second export read also removes repeated profile/catalog verification.

Validation

• RED 90a8350: deterministic SQLite WAL interleaving reproduces mixed descriptor, later-loss contamination, and validation schema race (3 failures).
• GREEN cbde051: all three pass; full local non-service suite 385 passed, 9 skipped, 53 deselected; focused checks and git diff --check pass.
• ca08100 corrects only the service test's native-state observation: PyMySQL caches an OK-packet status that SELECT EOF packets do not update. MySQL/MariaDB now record disabled autocommit honestly; the real concurrent-commit/old-and-fresh export semantics remain asserted. Reproduced the probe failure and pass on a disposable MySQL 8.4.9 instance.
• Independent full-diff local review at latest head: No findings.
• Latest-head push and PR CI both passed all 14 jobs: Python 3.11–3.14, package build/install, PostgreSQL 17.10/18.4, MySQL 8.4.11/9.7.2, MariaDB 11.4.12/11.8.8/12.3.2, and Dolt 2.2.2/2.2.3 interleaving. SELECT-only Dolt read/export tests also passed. No service claim is inferred from mocks.
• Existing Dolt propagation tests now observe the unchanged public read-profile gate rather than internal nested reader calls.

Boundary

Data/metadata consistency uses each engine's native snapshot guarantees. SQLite transactional schema rename is covered. This does not add writer/DDL coordination for non-MVCC server operations such as PostgreSQL TRUNCATE. Database transactions are not held during filesystem work.

Progress

Phase 2 #20 is merged. Next: phase 4 — bounded batched PHP server imports, in a separate PR.

Final review and handoff

• Full-diff local subagent review: No findings.
• @codex review completed for latest head ca08100, no findings and confirmed 👍.
• CodeRabbit completed: No actionable comments.
• All GitHub checks pass; merge state CLEAN. Three logical commits are pushed; worktree clean.

Phase 3 is ready for handoff. Do not merge automatically; the overall plan is not yet complete.