[haolpku]

Codex can read a spreadsheet through DataMind and then parse the same original again to calculate results or construct cell citations. This change lets the main Codex agent reuse versioned data directly in its analysis code.

• Keep DataMind optional, return compact operation summaries, and discover published sources/tables without opening SQL or embedding services.
• Share canonical table artifacts between SQLite snapshots and incremental Dolt builds. Preserve source row numbers, physical column mappings and conservative header suggestions; accept optional source-hash-bound field/range/unit mappings without changing raw values. Mapping-only changes reuse parsing and leave unchanged SQL tables intact.
• Add workspace_export_bundle: batch local JSONL/text files plus a source/version/cell manifest for subsequent Python and deliverable generation. Bound export size, reuse matching files, flag legacy lineage limitations and require a shared filesystem with the execution host.
• Keep pinned reads, publication, rollback and GC contracts; update the plugin, SDK guidance, packaged Skills and documentation. Existing direct live writes remain independent of Workspace snapshots.

Validation: local suite 482 passed, 91 environment-dependent checks skipped; wheel builds successfully. Real Dolt, migration, rollback and official SDK discovery were exercised in an isolated Linux image. Source-only export verification matched 1,141 tables / 253,577 rows / 2,556,071 cells against original or frozen source evidence and verified repeat-export reuse. GitHub CI supplies actual MySQL/Dolt and installed-wheel coverage.

The paired Workspace-Bench development pilot is still pending. Functional checks and exact-cell verification do not establish a task-quality or token-cost advantage. Keep this PR unmerged until the user reviews the experiment results. The coordinator still supports local single-host storage; remote execution must mount the export cache.