v2026.6.10

Fixed

• Migration journal reconciler — consolidated-DB self-heal. A cold open of a consolidated/sealed cleo.db could hang or fail with E_NOT_INITIALIZED ("Task database not initialized"). Root cause: probeAndMarkApplied ran its CREATE TABLE regex over raw migration SQL including prose comments, capturing phantom table targets (e.g. "…the project-side CREATE TABLE half of that move…") → the migration was never marked applied → Drizzle re-ran a bare CREATE TABLE against an existing table → threw (wrapped in .cause, so the "table already exists" retry guard missed it) → the rejected init-promise poisoned the connection cache → surfaced as a masked E_NOT_INITIALIZED. Fixes: (1) strip SQL line/block comments before DDL-target extraction (reusing the existing stripSqlComments idiom); (2) probe-tolerance for eliminated tables — a CREATE of a table a later migration permanently DROPs (e.g. release_manifests → superseded by releases) now counts as satisfied via a computeEliminatedTables() disposition walk; (3) zero-DDL migrations (pure-DML backfills, DROP-only, index/trigger-only) are gated by the consolidation cutover — pre-consolidation ones are stamped applied, post-consolidation ones are run-once by migrate() (so new backfills/drops actually execute); (4) retry predicates now walk the full err.cause chain; (5) a rejected init-promise is evicted from the connection cache so a transient open failure no longer poisons later callers. (provenance: T11829; #986)