fix: multi-database stress regressions, settings feedback, and grid alignment

[Blankll]

Complex SQL fixtures exposed crashes, lost session state, missing result sets, and incorrect metadata across PostgreSQL, MySQL, SQL Server, and Oracle. This fixes those adapter paths and makes docs/sql-test reproducible through the real Rust adapters and Java JDBC bridge. The same branch now also carries the settings-feedback pass and the result-grid fixes found while validating the fixtures in the desktop app.

Changes

Database adapters

• Preserve PostgreSQL multi-statement batches and render exotic types and arbitrary-precision values safely; correct index, composite foreign-key, identity, and generated-column metadata.
• Keep MySQL and SQL Server sessions across requests; preserve later and empty result sets, affected-row counts, and errors in later batch statements. Fix MySQL functional-index crashes and numeric/binary conversion, and SQL Server duplicate columns and missing precision.
• Preserve Oracle PL/SQL definitions and slash delimiters; fix JSON, timestamp, object-type, boolean, and structured-value conversion, and retrieve complete Oracle DDL with DBMS_METADATA.
• Keep quoted semicolons, comments, and PostgreSQL dollar-quoted bodies intact when extracting the editor statement at the cursor.
• Add four opt-in live regression suites, an SSH-aware runner, focused regression tests, and setup/coverage documentation.

Settings

• Add a shared SettingsFeedback component (pending/success/error with optional progress, status/alert roles) and use it across the settings sections.
• Providers: upsert by id, surface model-sync failures instead of swallowing them, and persist synced models.
• JRE/drivers, MCP bridge, plan and AI settings: consistent feedback states, retry affordances, and layout cleanup.
• Updater/download: reset and surface errors, and guard install behind the update check.
• DDL panel header reserves space for the dialog close button.

Result grid

• Numeric and boolean headers now take the same justification as their cells, so they line up with their values instead of the opposite edge of the column.
• Sort and filter indicators stay after the column name; they are pulled into the cell padding so the label itself stays flush with the values.
• Sort-priority numbers appear only for multi-column sorts.

Docs

• docs/sql-test/sample-queries.md: graduated easy-to-hard query ladder with per-engine dialect variants, an idempotent upsert per engine, and EXPLAIN examples.

Validation

• Live database verification: all four engines' schema/seed scripts and 84 query sections passed against the existing SSH test stack, including the documented expected-error cases.
• cargo test --manifest-path src-tauri/Cargo.toml: 547 tests passed; database-dependent tests remain ignored by default.
• npm test: 686 tests passed.
• Java bridge tests: 10 passed.
• npm run build: passed.
• npm run lint:check: zero errors; one pre-existing UnoCSS ordering warning in DefaultAvatar.vue.
• git diff --check: passed.
• Grid alignment measured in a rendered harness against the built stylesheet: numeric header and value right edges differ by 0 px (unsorted and single-sort); text headers stay left-aligned with 0 px offset.

Desktop visual interactions and complete DDL round trips are outside this automated coverage. The live runner recreates the documented disposable fixtures; Oracle's separate JDBC requests retain their existing autocommit behavior.