[bazer]

Problem

W0-F1 intermittently failed with SQLite error 1, cannot start a transaction within a transaction. Investigation reproduced the exact error using published Microsoft.Data.Sqlite 10.0.11 alone: two distinct, still-open connections receive the same pooled native handle.

The driver's activation method publishes its active flag before its weak owner. A concurrent checkout can misclassify that intermediate state as a leaked connection and lend it to another owner. Reversing the two writes fixes the reproduced interleaving without disabling pooling, adding locks, or retrying transactions.

Changes

• Record the cause, evidence, limitations and remaining dependency integration work.
• Preserve the standalone published-package reproducer and a tested upstream patch with regression coverage.
• Record a separate, hash-verified diagnostic archive; keep the original W0 evidence unchanged.
• Link the investigation from the 0.10 roadmap.

This PR changes no DataLinq runtime code or package references. The upstream patch is a reviewable artifact, not an installed dependency. Upstream submission is complete: issue dotnet/efcore#39008 and fix PR dotnet/efcore#39009. Both disclose that this fix was found, reproduced and posted with GPT-6 Astra on Extra High, with the repository owner's authorization. The contributor has personally posted CLA acceptance. Upstream review and a corrected package release remain pending. W0-F1 remains open.

The user has now explicitly approved a limited sequencing exception: internal W1 async contracts and controllable-provider tests may proceed on v0.10. SQLite integration acceptance, W0-F1 closeout, final provider feasibility/public API freeze and release approval remain blocked. No runtime workaround, dependency change or evidence relabeling is included.

This targets master because the defect affects current consumers. The eventual corrected-package adoption should land on master and merge forward to v0.10 under the accepted workflow. W0 capture is recorded separately in #144.

Validation

• Published 10.0.11 reproducer: duplicate live ownership and exact transaction error on wave 312.
• Controlled source interleaving: fails before the correction, passes after.
• Uninstrumented source stress: original fails on wave 351; corrected completes 800,000 opens with no duplicate ownership or other errors.
• New regression in the actual upstream test project: fails against original main, passes with the correction.
• Corrected upstream SQLite test application: 714 passed, 7 existing skips, 0 failed, with the normal upstream category=failing exclusion. This is not the full EF Core matrix.
• Corrected upstream project builds with 0 warnings/errors. Published/source diagnostic tests target .NET 10; upstream's actual test app targets .NET 11.
• Checked documentation links, byte identity of the embedded reproducer and patch, reverse application of the patch to the corrected upstream checkout, archive integrity and all 141 archived evidence file hashes.
• Markdown whitespace check passes; the preserved patch retains its required context-line whitespace. DocFX build is not applicable: docs/dev-plans/** is excluded from the website.

Raw evidence remains local; only the report, patch and standalone reproducer are committed. A passing stress run is not proof that all possible races are absent, and no passing run retroactively changes the original W0 failure.