refactor(ios): streaming data architecture for large tables and queries

[datlechin]

Summary

Rebuilds the iOS data layer around streaming so opening large tables and running ad-hoc queries no longer materialises the entire result set in memory. Eliminates several latent OOM crashes reported on TestFlight (the most recent being the 100k-row table bug fixed in #1041; same shape would still have killed the app on a single 700 MB BLOB column).

Approach

• New shared types in TableProModels (used by both iOS and macOS). Cell enum has four cases: .null, .text(String), .truncatedText(prefix:totalBytes:ref:) for text > 4 KB, .binary(byteCount:ref:) for any binary column. Wide cells carry only metadata plus a CellRef (table + PK values) for re-fetching the full value on demand. Row wraps [Cell]. StreamElement is the streaming event type. StreamOptions carries truncation thresholds and lazy context.
• New executeStreaming(query:options:) method on DatabaseDriver with a backwards-compatible default implementation that bridges through legacy execute(query:). macOS app and all 14 plugins keep working unchanged through the default; iOS drivers override with real C-level streaming.
• MySQL: switched from mysql_store_result (buffered, libmariadb mallocs the full result before Swift sees the first row) to mysql_use_result (server-side cursor, one row per fetch).
• Postgres: switched from PQexec to PQsendQuery + PQsetSingleRowMode + PQgetResult per-row.
• SQLite: keeps sqlite3_step row loop. Stops eagerly base64-encoding BLOBs; emits Cell.binary(byteCount:ref:) so the bytes never enter Swift memory.
• Cancellation: AsyncThrowingStream.onTermination propagates to the C layer (PQcancel, sqlite3_interrupt).
• MemoryPressureMonitor (new actor): wraps DispatchSourceMemoryPressure + os_proc_available_memory; started at app launch.
• RowWindow (new): bounded sliding buffer (default 200 rows) replacing the unbounded [[String?]] @State arrays in views.
• DataBrowserViewModel / QueryEditorViewModel (new, @Observable): own the RowWindow, drive the streaming, listen to memory pressure, call RowWindow.shrink(to:) on warning (100 rows) or critical (50 rows + cancel).
• StreamingExporter (new): writes CSV / JSON / SQL to a FileHandle in O(1) memory regardless of result size.
• RowDetailView: TabView (which pre-rendered 2-3 full rows on iPad split view) replaced with single rowContent + horizontal DragGesture for prev/next. Memory cost O(1) per row.

Test plan

Run on a real iPhone (not Simulator) to exercise iOS memory enforcement.

• Compile clean: iOS app builds on Debug + Release. macOS app builds (no protocol regression).
• MySQL streaming: open a 1M-row table on a 4 GB iPhone. Should stream first page in under 1 s, no Jetsam kill.
• Wide cell: row with a 50 MB TEXT column. Cell shows truncated prefix + total byte count, no OOM.
• BLOB: row with a 200 MB BLOB. Cell shows [BLOB 200 MB], no decode at fetch time.
• Cancellation: in editor, run SELECT * FROM big_table. Tap Stop. Verify PQcancel / sqlite3_interrupt / mysql_use_result cleanup actually fires.
• Memory pressure: Xcode Debug -> Simulate Memory Warning. DataBrowserViewModel.window should shrink and post a toast.
• iPad split view: open browser in slide-over alongside Notes. RowDetailView no longer pre-renders adjacent rows.
• Streaming export: 100k-row CSV via share sheet. Memory profile flat (~100 MB) instead of spiking to file size.
• macOS regression: xcodebuild test -scheme TablePro -only-testing:TableProTests/MySQLCreateTableTests still passes (no behaviour change in macOS plugin path).
• swiftlint lint --strict clean on every new file.
• Banned-words / em-dash sweep clean.

What's NOT in scope (deferred)

• iPad NavigationSplitView regular-size-class adaptation in ConnectionListView. The OOM root cause is fixed by streaming + windowing; the iPad split view is UX polish for a follow-up.
• macOS app and plugins adopting executeStreaming. They keep using the legacy execute(query:) path via the protocol default. A future PR can switch them piecemeal.
• Full Cell-aware UI (per-cell "Load full value" button in RowDetailView). The current legacyValues bridge surfaces wide cells with placeholder strings; the lazy-load button can ship in a follow-up that doesn't change the streaming layer.

Risk

• Big diff (~1300 LOC). Mitigated by: backwards-compatible protocol default keeps macOS/plugins green; iOS view layer keeps existing UI shape and just rebrands the data source.
• mysql_use_result semantics: requires draining all rows before the next query. The actor's endStream does this. Cancellation drains too.
• PQsetSingleRowMode: must be called immediately after PQsendQuery, before any PQgetResult. Implementation matches that order.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.