# v0.105.3: DVM corrections and release publication safety

> **Status:** Released
> **Scope:** Patch
> **User promise:** "Known wrong-result cases are fixed, and every release channel waits for qualification."
> **Blocked by:** [v0.105.2](v0.105.2.md)
> **Series:** [v0.105.x](v0.105.x.md)
> **Source:** [Post-v0.105.2 assessment](../pg-trickle-v0.105.2-assessment.md), sections 3 and 4.1

## Outcome

Ship the DVM corrections already merged in PR #1011 with compact regressions
and guidance for affected existing results. Close the Docker publication path
that can bypass qualification. These known defects justify a maintenance
patch under the v0.105 series policy.

## Work

- Establish the affected release range with minimized reproductions. Preserve
  the aggregate predicate correction and the unchanged-left relation used by
  semi-join and anti-join maintenance.
- Extend the existing exact-result tests with simultaneous changes to both
  sources of `EXISTS` and `NOT EXISTS`. Cover deleted left rows, first and last
  right matches, duplicates, nullable keys, and a downstream aggregate.
- Add a regression for a filter above a projected subquery whose reconstructed
  SQL contains an inner `WHERE`. Assert both valid SQL and the expected result.
- Show that each reproducer fails on the affected implementation and passes
  on the candidate. Assert bag equality and the intended effective strategy
  after each mutation cycle. Check output-delta multiplicity when publication
  is enabled. A whole-query `FULL` refresh cannot satisfy a differential test.
- Document how to identify potentially affected stream tables, compare their
  results with the defining query at a consistent boundary, and repair or
  recreate affected state using supported procedures. Include downstream
  consumer invalidation and resnapshot where needed. Installing the fix alone
  does not prove that existing materialized results are correct.
- Make every official publication path depend on successful qualification,
  including per-architecture containers, manifests, version tags, and mutable
  release aliases. Test the workflow dependency graph and a deliberately
  failing qualification without promoting public artifacts.
- Requalify the exact packaged candidate and retain the regression and
  publication-gate results. State which checks actually ran for each artifact.

## Exit criteria

- [ ] Reproductions identify the affected versions and demonstrate failure
      before the correction and success after it.
- [ ] Exact rows, duplicate counts, strategy, and applicable output deltas pass
      on the packaged candidate across the named mutation cases.
- [ ] A tested recovery procedure repairs affected state without requiring
      every stream table to be rebuilt or losing committed source changes.
- [ ] Failed, skipped, or cancelled required qualification prevents all
      official release promotions, including container aliases.
- [ ] The patch publishes candidate-specific qualification evidence and
      affected-query guidance. Unexecuted checks remain explicitly unqualified.

## Existing implementation points

Use [aggregate SQL generation](../src/dvm/operators/aggregate.rs),
[semi-join maintenance](../src/dvm/operators/semi_join.rs),
[anti-join maintenance](../src/dvm/operators/anti_join.rs), and the
[differential equivalence suite](../tests/e2e_diff_full_equivalence_tests.rs).
Update the existing [release workflow](../.github/workflows/release.yml) and
[upgrade guidance](../docs/UPGRADING.md).

## Scope control

No new query families, public contracts, capture modes, or controller actions.
Keep Graph V1 and Delta V1 unchanged. The broader evidence and performance
work in [v0.106.0](v0.106.0.md) does not delay this corrective patch. The
72-hour soak, longevity environment, and v1.0 qualification remain deferred.