progress.md

Syntax audit progress

Data transfer — 2026-10-19

• Began authorized implementation; read root instructions and native macOS/SwiftUI/concurrency/planning skills. Preserving earlier planning records and untracked metadata.
• Located and inspected local DBX UI, task store and actual Tauri/core transfer runner. No GitHub search needed for ongoing implementation.
• Inspecting QueryCraft session/schema/connection APIs before choosing the smallest shared execution integration.

Code commit — 2026-10-14

• User explicitly requested a code commit. Reviewed release instructions and fetched tags; current published baseline v0.1.9 / API 5 remains unchanged. Six driver source/package manifests unchanged, and all 12 public architecture catalogs contain compatible API 5 packages.
• Committed 69 editor feature/test/grammar/design/smoke files as dab2db0 on main. Staged whitespace check passed after removing a trailing blank line from a new test. Kept local planning records and generated Xcode Package.resolved untracked. No push, tag or release performed.

Built-in readability correction — 2026-10-14

• User asks that built-in presets themselves pass readability, not merely explain their failures. All six previous palettes failed selection contrast; lowest selected comment was around 2.9:1.
• Deepened light accents without changing hue families, reduced dark selection brightness and brightened weak dark accents/comments. Added unfocused grayscale selection to the existing advisory calculation. Strengthened the preset regression from normal/current-line backgrounds to every result at unchanged thresholds.
• All 18 methods in color-settings/customization suites passed, including six preset/mode cases requiring every advisory result to pass. Native selection/search rendering and foreground pixel regressions passed; six palette screenshots inspected. No custom overrides are rewritten; existing pure presets resolve the updated built-in colors automatically. Main Debug build in progress.
• Debug build passed; signed Dev installed with marker 20261014-143528-53e23f06 and previous app backed up. git diff --check passed. Logs: /tmp/querycraft-editor-readable-presets-{tests,build,install,launch}.log.

Layout feedback — 2026-10-14

• User prefers the previous side-by-side layout and asks what readability ratios mean. The 600-point sheet minimum overrode its 840-point ideal, causing the vertical fallback to appear by default.

• Restored persistent left preview/right scrollable settings, a 420-point preview and explicit 960×700 minimum. Readability now explains selected-text-only cases, keeps details collapsed, labels current versus suggested ratios and aligns entries left.

• Running layout/presentation regression and rendering both languages and appearances before rebuilding Dev.

• Initial new layout assertions used fittingSize on a host with sizingOptions disabled (which correctly returns zero) and assumed a 500-point editor height despite English guidance wrapping. Removed the invalid fitting measurement (covered by the separate sizing-enabled presentation test) and assert >450-point height, still excluding the former 220-point vertical preview.

• Layout and presentation suites passed: six methods, including two-width/light-dark/Chinese-English rendering, minimum sheet size, existing editor selection/undo and JSON capture checks. Inspected 960-point dark Chinese and light English and 1080-point light Chinese screenshots. Main Debug build running.

• Debug build succeeded; signed Dev installed as 20261014-142802-5f128345, with previous install backed up. Layout regression log: /tmp/querycraft-editor-layout-regression2.log; build/install/launch logs use /tmp/querycraft-editor-layout-*.log. No driver/execution changes or desktop input simulation.

• Read SQL/JSON capture rules, index recovery, parser and existing tests.

• Building a reproducible syntax corpus, including incomplete input and incremental edits.

• Added 30 SQL fixtures, 8 shared JSON fixtures and 7 Elasticsearch fixtures.

• Added specific capture rules and display-only numeric/Doris recovery, sharing opaque quoted-token scanning with index recovery.

• First compile found an invalid raw-string escape for #; removed the unnecessary regex escape before rerunning.

• Expanded SQL corpus to 38 cases. Four incremental append/delete streams matched fresh parsing; partial-range comparisons passed. One regression in JSON arrow operators caused by splitting a minus was corrected by preserving multi-character arrows in the scanner.

• Regenerated JSON parser from retained upstream grammar with Unicode escape/exponent corrections.

• Running SQL structural/color, JSON coverage and Elasticsearch console regressions together.

• SQL corpus, incremental streams, partial highlights and existing structural tests passed. JSON exponent and full Unicode escapes passed after regeneration.

• Multi-request Elasticsearch still exposed loss of earlier body captures because request lines were parsed as JSON. Replaced this with cached independent fragment clients and stale-revision rejection, retaining normal client parsing for shared JSON views.

• 55 syntax fixtures (38 SQL, 9 JSON, 8 Elasticsearch), four per-character append/delete streams and stale-source rejection pass. Main Debug build passed.

• Final broad suite found an outdated JSON key expectation in the pre-existing viewer test; corrected type to property and rerunning the full selected suite.

• Final verification: 83 selected tests passed; Debug build succeeded; git diff --check passed. Signed Dev installation succeeded with backup. No commit, tag, release or live database/request execution.

• Logs: /tmp/querycraft-syntax-verified-tests.log, /tmp/querycraft-syntax-final-build.log, /tmp/querycraft-syntax-install.log.

Expanded coverage follow-up

• Expanded to 130 annotated fixtures (78 SQL, 28 JSON, 24 Elasticsearch), including generated/check columns, CHANGE COLUMN, schemas, materialized views, lateral queries, tuples, JSON paths, escaped identifiers, typed literals, root JSON scalars, every JSON escape, surrogate pairs, duplicate keys, CR/CRLF, request comments and NDJSON.
• Added 216 SQL middle replacement/undo transitions, JSON/request middle edits and clear/restore checks, 300-statement SQL and 180-body JSON/Elasticsearch viewport tests.
• First expanded pass: all JSON/request and edit/viewport checks passed; SQL exposed missing CHANGE COLUMN old-name, CREATE SCHEMA, typed DATE prefix and escaped quoted-name captures.
• Added capture rules and corrected doubled quote/backtick grammar tokens. Added three structural checks ensuring semicolons and SQL-looking text inside escaped names/strings do not create executable statements. Regenerating pinned SQL parser before broader regression.
• Regeneration completed; generated parser diff contains only two lexer transitions. All 207 tests in 15 suites passed, including 130 fixtures, SQL formatting/completion/execution safety, both appearances and JSON rendering. Debug build passed. Installing verified Dev build.
• Verification logs: /tmp/querycraft-expanded-regression-tests.log and /tmp/querycraft-expanded-build.log. Detailed suite log: /Users/weilai/Library/Developer/XcodeBuildMCP/workspaces/QueryCraftPublic-21b5c5ce48ed/logs/swift_package_test_2026-10-14T02-25-01-114Z_pid58844_078cacd9.log.
• Signed Dev installed and launched successfully; install marker 20261014-102601-7dcfdd78, backup retained by installer. git diff --check passed. Removed only the build-generated untracked Xcode Package.resolved; no commit/tag/release or desktop input simulation.

Full customization implementation

• Added palette history/session restore, strict versioned import/export, persisted personal themes, HEX fields, native color wells with fixed appearance/role ownership, readability checks, typography and five extra surface roles.

• Added adaptive settings and five editable previews with capture-to-role navigation and search preview.

• Added isolated dialect display tree, query connection ANSI_QUOTES preference and context propagation through DDL; original execution parse is unchanged.

• First compile caught associated-value emphasis style and optional emphasis manager usage; fixed recoloring to affect only standard search layers and preserve explicit bracket colors.

• Customization/model/native-well/dialect tests passed; fixed legacy layout test mismatch by retaining an 840-point ideal width with adaptive minimum size. Multi-editor sync and two-language 600/840-point layout checks passed.

• MB tests exposed quadratic SQL scope/reference scanning; scoped references with sorted-offset lookup. Further profiling found uppercase scientific notation triggers grammar error recovery despite display-only recovery; fixing the retained grammar and regenerating. Added bounded parser recovery and fresh-parse fallback instead of allowing unbounded incremental work.

• MB JSON and ES measurements already passed: ~1.02 MB JSON setup 3.42 s / edit 26 ms / 3 viewports 6 ms; ~1.15 MB ES setup 5.20 s / edit 409 ms / 3 viewports 864 ms (Debug). SQL requires another performance pass after grammar regeneration.

• ES viewport profiling exposed canonical Unicode String comparison and repeated request-line scans. Switched to literal UTF-16 identity and cached request captures; latest ES viewports ~149 ms, JSON ~8 ms. Added canonical-equivalent replacement regression.

• SQL profiling also exposed redundant syntax snapshots and indexed tree-sitter child enumeration. Reused the initial snapshot, skipped unnecessary diagnostic/delete/update traversals and switched child enumeration to cursors; performance validation remains in progress.

• Final functional regression: 310 tests in 23 suites passed (2026-10-14 11:49); includes existing execution, formatting, completion/snippet and query-document tests. A native test-window over-release was corrected by setting isReleasedWhenClosed=false for Swift-owned windows used by asynchronous completion tests; focused 44-test rerun and full rerun passed.

• Final isolated MB performance suite passed without relaxing budgets: SQL 1.134 MB initial 8.925 s, edit 2.883 s, 3 viewports 57.8 ms; JSON 1.018 MB setup 2.775 s, edit 27.2 ms, viewports 5.2 ms; ES 1.269 MB with 5,500 request comments setup 7.687 s, edit 383 ms, viewports 93.5 ms. This is Debug on the local machine. Peak RSS growth remained below 768 MiB; later-case high-water deltas are not independent per-document memory measurements.

• Verified incremental statement/AST/scope reuse against fresh parses, including changed lengths, Unicode, semicolon split/merge, nested CTEs and UPDATE/DELETE WHERE changes. Display recovery uses a local statement scan only when reliable old/new boundaries match, otherwise retains full scanning.

• Settings screenshots now render SwiftUI layers plus native editor content correctly; inspected light/dark, Chinese/English and 600/840 widths. Six palette render cases exercise real search layers and recolor existing active/inactive highlights. Main app Debug build is running.

• Completed: Debug build succeeded, six-palette active/inactive background and text-layer update assertions passed, git diff --check clean. Signed Dev installed and launched, marker 20261014-115535-ef949ad5; previous Dev retained in installer backup. Removed only generated untracked Xcode Package.resolved. No commit/tag/release or live database/HTTP validation.

Final regression continuation — 2026-10-14 afternoon

• Fixed search foreground painter order and verified rendered pixels.

• Added real theme-file cancellation/error/duplicate/persistence coverage, native color-panel target/undo coverage, five-provider composition coverage, and executed-SQL typography coverage.

• Fixed a reproducible asynchronous stale-highlight crash found by the full suite; added three deterministic regression methods for request/edit generations.

• 318 functional checks verified (317 in the combined run; the lone load-sensitive timing case passed isolated at 2.544s without threshold changes). The separate MB performance suite passed.

• Final Debug build passed; all six installed Driver API 5 drivers passed read-only local-service smoke. Rendered 18 native SQL/JSON snapshots. Actual MySQL and Doris DDL exposed function-index/foreign-key and parameterized-type gaps, now fixed with four added corpus fixtures (134 total) and focused regression passing.

• Search foreground now passes light/dark rendered-pixel assertions, including absence of the original syntax glyph underneath. Final Debug build was signed and installed as Dev marker 20261014-135801-41722e67; installer retained the previous app backup.

• Physical desktop input remains unverified because the required reliable input isolation is unavailable: no real mouse drag, file-dialog clicking, or system IME candidate-window simulation was performed.

• Added typed transfer options/driver capability matrix, atomic task/folder persistence with cycle/name validation, token-aware SQL rewriting and binary-safe conversion, catalog readers for relational objects, a metadata planner with FK/index/ownership handling, and transaction-aware background execution.

• Execution design: source repeatable-read snapshot (Doris explicitly requires quiescence), complete HEX binary reads, deterministic paged ordering, per-table transactions for append/upsert/rebuild; overwrite uses one data transaction across selected tables to preserve rollback with FK-linked targets. Rebuild backups are only cleaned after all selected operations succeed.

• Starting compilation before the UI integration to catch protocol/concurrency mismatches early. No Driver API or external driver changes yet.

• First focused run passed all 19 Swift Testing methods (including parameterized strategies/filters) across SQL safety/value conversion, task library and transaction runner suites. Latest subsequent metadata/UI refinements still require rerun.

• SwiftPM test wrapper attempted network dependency updates; stopped that run and used XcodeBuildMCP macOS tests with cached pinned packages and explicit only-testing arguments instead.

• One patch attempt was rejected because it contained an empty update hunk; no partial changes were applied, and the corrected patch succeeded.

• Added a live fixture runner for MySQL 5.7→8.0, PostgreSQL 17→17 and both MySQL/PostgreSQL cross-engine directions. It uses UUID databases/schemas, in-memory credentials and cleanup, never saved user profiles.

• Second focused run: 21 test methods in four suites passed (0 failures/skips), including native offscreen workbench layout at 1120×790 and 920×680 in both appearances, error/history transitions and dirty task switching.

• Live tests passed MySQL 5.7→8.0, PostgreSQL 17→17, MySQL→PostgreSQL and PostgreSQL→MySQL. Each copied 203 rows with complete non-UTF8 binary bytes, 30-digit decimals, Unicode/quotes/backslashes, NULL and empty values; append/create, Upsert, overwrite, filtered overwrite and cancelled-overwrite restoration all passed. Same-engine cases also passed FK-related rebuilds, view replacement, secondary-index creation and original backup cleanup.

• The standalone live compiler initially lacked transitive C shim module maps. Replaced the static assumption with discovery/deduplication of public source module maps; the complete live matrix then passed. Temporary PostgreSQL container and UUID fixtures were cleaned.

• Added PostgreSQL enum/enum-array preparation, identity ALWAYS preservation and explicit column collation/table comments after these test builds; validate these refinements in the next run.

• The new deterministic slow-request fixture initially waited for the selected database, but catalog database enumeration deliberately connects without selecting one. This left the test gate unopened and stalled the run. Changed the gate to pause its first request independently of database selection, stopped only that test command's process tree, and will rerun the loading tests.

• Expanded live matrix also passed PostgreSQL enums/enum arrays/defaults, GENERATED ALWAYS identity, stored generated columns, standalone sequence position and target routine execution. All six installed API 5 drivers loaded/activated against the new host framework.

• First guarded capture invocation filtered a Swift Testing method ID without parentheses, causing zero tests to execute; the supervisor now selects the entire presentation suite and requires four real capture acknowledgements, so a zero-test run cannot be reported as visual acceptance.

• Second guarded UI run was explicitly stopped by user Escape before screenshots. Supervisor stopped its test process group and removed the shield. No further desktop UI operations will be performed in this session without renewed authorization. Offscreen layout tests are still valid, but their cached bitmaps omit system-control layers and are not visual acceptance.

• Added canonical MySQL DDL escaping: same-engine SHOW CREATE uses captured source NO_BACKSLASH_ESCAPES semantics; generic cross-engine MySQL DDL uses NO_BACKSLASH_ESCAPES, with strict mode verified. This preserves quotes/backslashes in column comments and defaults while keeping row literals mode-independent. Latest compile/tests/live matrix rerun pending.

• Latest Debug application build passed. The SQL-mode revision also passed all 84 selected methods before the final view/trigger metadata refinements.

• PostgreSQL copied views now retain reloptions and CHECK OPTION; materialized views and new triggers are installed after data. PostgreSQL triggers on rebuilt tables are created anew without dropping the trigger retained on the original backup table.

• Routine matching now uses canonical argument types (with namespace-aware token comparison) rather than parameter display names, while preserving the full signature for UI and source OIDs for catalog lookups.

• Final verification in progress; visible UI testing remains stopped at the user's Escape request.

• Final application build succeeded after all source changes. Final 84-method regression passed (0 failures/skips); final expanded live matrix passed all four directions, including uppercase targets, complete SELECT filters and custom timeout 0/1.

• Real PostgreSQL checks additionally assert enum/array values, column comments with quotes/backslashes/Unicode, view security_barrier/security_invoker/CHECK OPTION, materialized view population after data, trigger installation/rebuild, standalone sequence position and identity's next generated value after explicit inserts.

• Installing the signed isolated Dev bundle in the background. Installation wrapper blocks any osascript attempt to quit an unexpectedly running Dev instance; no window launch or further desktop interaction after Escape.

• Signed Dev installation succeeded, marker 20261019-155653-d6bcfc35. Verified bundle ID io.github.future0923.QueryCraft.Dev; previous bundle retained in .build/dev-app-backups/20261019-155653-d6bcfc35. No app launch, commit, tag or release. Temporary PostgreSQL container and transfer databases cleaned; original two MySQL fixture containers retained.

• Delivery documentation now records final 84-method regression, expanded live matrix, six-driver loading and the incomplete visible UI acceptance after user Escape. Final logs: /tmp/querycraft-transfer-release-check.log, /tmp/querycraft-transfer-final-build.log, /tmp/querycraft-transfer-final-live.log, /tmp/querycraft-transfer-install.log.

Transfer workbench layout feedback — 2026-10-19

• User supplied a screenshot showing a 320-point empty task sidebar, a fixed-height endpoint gap, configuration below the scroll fold and an empty 64-point status row. Rebuilt the layout with a 208-point task column, compact 112-point endpoint cards, flexible object list, always-visible basic options and a single 56-point footer.
• Preserved transfer configuration and execution behavior. Moved per-table filters into selected table rows, bulk actions into the object menu, batch/timeout/ownership into an advanced popover, and task name/folder into a save sheet with local drafts. Initial and empty states provide specific guidance without replacing containers during requests.
• First added accessibility-tree geometry check was unavailable for hidden native windows and was removed. Existing plus initial-layout regressions passed: 5 methods, 0 failures/skips. CALayer offscreen rendering includes the native controls omitted by the earlier cacheDisplay snapshots; inspected light/dark 920-point empty and selected states. No simulated mouse/keyboard or foreground screenshot validation. Final app build and Dev installation follow.
• Final Debug build passed. Signed Dev updated and reopened, marker 20261019-161457-f6c8802a; previous app backed up. Logs: /tmp/querycraft-transfer-redesign-tests.log, /tmp/querycraft-transfer-redesign-build.log, /tmp/querycraft-transfer-redesign-install.log. No driver or transfer-execution changes; git diff --check passed. No commit, tag or release.

Transfer workbench hierarchy feedback — 2026-10-19

• User found the revised page visually scattered. Replaced the two endpoint cards and bottom options strip with aligned source/target rows above a unified object-list/configuration split. Removed repeated empty-state icons/counts, matched option menu widths and input styling, shortened strategy labels with contextual explanations, and collapsed the task sidebar by default when the library is empty. Saved tasks remain available from the labelled toolbar toggle, including from Progress.
• Initial focused 5-method regression and Debug build passed. Inspected light empty state without sidebar and dark populated state with sidebar at 920-point width. Rechecking the final full-width menu adjustment before signed Dev update. No mouse/keyboard simulation or new foreground UI verification.
• Native SwiftUI menus did not honor the intended full-width labels in offscreen inspection, so replaced the three option controls with standard NSPopUpButton representables. They retain disabled strategy options, accessibility labels, selected values and model callbacks. Added actual NSView width/alignment/bounds checks and selection-to-model verification to the existing presentation tests. All 5 methods passed, 0 failures/skips; final native snapshots inspected. Final app build/install follows.
• Final Debug build passed; signed Dev installed and reopened, marker 20261019-163229-d1a53bba, with previous bundle backed up. Final logs: /tmp/querycraft-transfer-layout-v3-native-tests.log, /tmp/querycraft-transfer-layout-v3-native-build.log, /tmp/querycraft-transfer-layout-v3-install.log. git diff --check passed. No changes to transfer execution or drivers; no commit, tag or release.

Transfer workbench detailed feedback — 2026-10-19

• User prefers the existing grid-field search picker, requests removing the transfer safety lock, removing the Saved tasks focus rectangle, improving faint text, and showing batch/timeout directly. Extracted a shared native search/list surface from WorkspaceGridSearchColumnPicker and reused it for transfer locations, preserving search icon, opaque background, left checkmark, row selection and keyboard navigation.
• Removed the transfer-only lock from manager guards, window initialization, footer and review sheet; other workspace locks are unaffected. SQL preview, destructive-scope acknowledgement, target reservations and metadata checks remain. Batch size/timeout and applicable ownership are inline; guidance/explanations now have stronger text contrast. Saved tasks uses a native unbordered NSButton with focusRingType=none and retains keyboard activation.
• Added direct reviewed-plan execution regression, focus-ring assertions and native search-surface snapshots; running transfer presentation/runner and existing grid-search regressions before final build/install. No real mouse/keyboard input or new foreground visual automation.
• All 26 presentation/runner/grid-search methods passed. Inspected native search picker snapshots and the smallest empty workbench: opaque list, search icon, left checkmark and whole-row highlight match the existing field selector; inline batch/timeout and darker guidance fit without a popup, and the transfer lock is absent.
• User additionally requests database names in the SQL review object mapping. Added individually quoted source/target database-qualified names, including PostgreSQL schema, preserved exact target-name case, escaped newlines in comment labels, and removed trailing dots from the preview location heading. Added MySQL/PostgreSQL/Doris mapping regression cases; running the final combined set.
• Final combined regression passed: 36 methods, 0 failures/skips, covering transfer SQL/runner/presentation and the shared grid picker. Final log: /tmp/querycraft-transfer-polish-final-tests.log. Building the app before signed Dev replacement; no further source changes planned.
• Final Debug build passed. Initial installation stopped because the running app returned AppleEvent user-cancelled (-128), leaving the verified staged bundle intact. User explicitly approved quitting/updating; normal quit then succeeded. Reverified and installed that signed staged bundle, marker 20261019-164722-ddfa9b54, backed up the previous Dev and reopened it. No forced process termination or desktop input simulation. Build/install logs: /tmp/querycraft-transfer-polish-build.log and /tmp/querycraft-transfer-polish-install.log (initial stopped attempt). No commit/tag/release.

Transfer list separators and review summary — 2026-10-19

• User requests fewer underlines in the object list. Hidden native row separators and replaced native Section headers with plain labelled header rows, because native macOS grouping still drew a line despite section separator visibility. Kept object rows, selection/filter controls and region boundaries intact.
• User asks what the two identical preview comments mean: they were one batch-transfer annotation per selected table, detached from the table mapping lines. Consolidated into one explicitly counted table-transfer summary with strategy and batch size; all qualified object mappings remain. Extended the existing mapping regression to two tables and require a single summary.
• First 6-method presentation run passed; final SQL/runner/presentation rerun in progress. Native screenshots inspected; no desktop input simulation or foreground UI validation.
• Final SQL/runner/presentation regression passed: 26 methods, 0 failures/skips. Final plain-header display checks passed in a further 6-method presentation run; inspected the dark populated list to confirm the native group underline is gone. Debug build passed; signed Dev installed/reopened with marker 20261019-165820-25888978 and previous bundle backed up. Logs: /tmp/querycraft-transfer-clean-preview-tests.log, /tmp/querycraft-transfer-separators-final-layout.log, /tmp/querycraft-transfer-clean-preview-build.log, /tmp/querycraft-transfer-clean-preview-install.log. git diff --check passed; no commit/tag/release.

Transfer object search width — 2026-10-19

• User requests that object search fill the header's unused space and removes the lower unused horizontal line. Removed the header spacer/fixed 130-point search width, kept the title/menu at intrinsic widths and made search flexible. Removed the divider above the fixed-height object-count/loading footer; content and loading layouts remain stable. Running existing presentation regression before build/install; no new tests for this cosmetic change.
• Resumed on 2026-10-20 after the interrupted test log ended before results. Existing 6-method presentation suite passed; inspected the dark minimum-width snapshot, then Debug build and signed Dev installation succeeded. User additionally requested removing the Objects heading, so removed it and am validating the final search-only header before replacement/reopen. No desktop input simulation.
• Final search-only header passed all 6 presentation methods (0 failures/skips); inspected the updated dark minimum-width render. Test build included the final app. Signed Dev installed and opened with marker 20261020-075609-a5149971, prior bundle backed up. Logs: /tmp/querycraft-transfer-search-header-tests.log and /tmp/querycraft-transfer-search-header-install.log. No transfer execution or driver changes; no commit/tag/release.

Large live Docker transfers — 2026-10-20

• User requests that I perform the large-data test. Started Docker and added a reproducible benchmark using the current app framework, installed MySQL/PostgreSQL drivers, actual planner/manager/runner, uniquely named isolated containers and synthetic data. No saved connection profiles or real databases are used, and no desktop input is simulated.
• Initial scope: 1,000,000 and 10,000,000 rows across MySQL 5.7→8.0, PostgreSQL 17→17, and both cross-engine directions, default 1000-row batches. Monitor process RSS and main-actor tick gaps, close the workbench model while transfer continues, verify persisted history, exact row count/ID sum and every payload/decimal/binary/boolean value. Append and overwrite cancellation must roll back; overwrite must preserve an existing sentinel row. Current results directory is recorded in .build/transfer-large-run.log.
• Baseline inspection found byte-by-byte String(format:) dominating MySQL literal encoding (100k rows in 48.37s; 200k in 103.22s). Retained a 2-second process sample and progress under .build/validation/transfer-performance/20261020-082430-a2b606; intentionally stopped that incomplete baseline and removed its isolated fixtures before rebuilding. Replaced per-byte Foundation formatting with direct hex digit encoding and added guarded exact-numeric single-primary-key pagination; full SELECT, composite/unsupported/generated/no PK and Doris retain OFFSET. No driver/API changes.
• All 28 transfer SQL/runner/table-export regression methods passed. Tests cover exact IDs above 2^53, filter composition, cursor injection rejection, fallback modes, batch boundaries, rollback/cancellation and MySQL/Doris byte-preserving text encoding. Final full matrix runs from .build/validation/transfer-performance/20261020-082925-b71930; initial 100k now takes 11.06s. Full results are pending, not yet counted as passed.
• All four 1,000,000-row cases completed and passed full deterministic row validation: MySQL→MySQL 115.41s/49.52MiB peak RSS, PG→PG 94.23s/44.55MiB, MySQL→PG 117.93s/45.74MiB, PG→MySQL 105.07s/52.02MiB. IDs start above 2^53, each target has correct row count/ID sum and zero invalid payload/decimal/raw/flag values. History persistence and workbench-model close path also passed. These are Debug engine harness RSS measurements, not the visible app's total memory.
• Both sources now have 10,000,000 rows. Full larger matrix is running; do not claim it passed yet. Saved query-plans.txt confirms MySQL range access on PRIMARY and PostgreSQL Index Scan with exact BIGINT cursor literals. Current running session is 8852 and log is .build/transfer-large-final.log; three qc-large-20261020-082925-b71930-* containers belong to this test only and the script removes them on completion/interruption.
• User asks about more speed improvement. Current sample shows remaining client-side INSERT construction/type and string work; harness around 80% of one CPU core, target MySQL about 10%. Answered in commentary with measured 8.5–10.6k rows/s at 1M, and priorities: prepared column encoders/string assembly, batch-size comparisons, Release comparison, then native COPY/bulk pathways (larger driver scope). No additional implementation promised or applied; the current 10M matrix continues with frozen source. Optimized sample: mysql-optimized.sample.txt. Source table sizes: MySQL 1,936,719,872 bytes; PG 2,230,296,576 bytes. MySQL 10M run has passed 5M rows around 52.25MiB peak RSS; not yet completed.
• Added an isolated compilation comparison of the exact current TransferSQL/WorkspaceSQLExportDialect sources (copies plus testable host import in .build only). 100k rows / 35,862,200 SQL bytes: -Onone 9.09785s, -O 3.20267s (~2.84× encoding-only improvement). This is not an end-to-end Release app benchmark; told the user that distinction. Results and source copies in encoding-comparison/, helper .build/compare-transfer-encoding.py. An initial optimized compile needed complete Sparkle artifact headers; fixed framework search path without touching app products. Actual source query digests saved to actual-query-digest.txt confirm emitted keyset SQL (8819 queries totaled ~24.7s in source DB at sampling time). MySQL 10M run is around 8.55M/52.25MiB; not complete yet.
• MySQL 5.7→8.0 10M completed: 1221.24369s, 8188.37 rows/s, 52.25MiB sampled peak RSS; all 10M target rows verified with correct ID sum 50,000,005,000,000 and zero invalid values. Background manager/model-close and persisted final history passed. Main-actor 100ms heartbeat max gap was 1.825s in this run (which also included CPU sampling and the separate encoding compile/run); do not claim a perfectly smooth visible UI or attribute the gap without further evidence. PostgreSQL 10M and remaining cross-engine/cancellation cases are still pending. Final source remains frozen; app not updated with these performance changes yet.
• User explicitly steered: optimize first, use Release for the 10M test, avoid wasting time on the Debug matrix. Acknowledged the oversized Debug plan and requested cancellation through the harness control file. PG stopped at 1,606,000 read rows in 0.106s, reported cancelled/0 committed; the script then removed all three owned containers. This intentionally incomplete Debug matrix is a retained baseline, not a failed production regression. Do not rerun the remaining Debug large cases.
• Second optimization: prepared per-table INSERT/cell encoders cache field mapping, type conversion, identifier/identity/upsert SQL; append rows directly into one string and accept slices to avoid row/tuple arrays. UTF-8 byte escaping replaces repeated Foundation replacements/Character scans, retaining MySQL/Doris hex mode, PG E-string/NUL rules, exact binary and boolean conversions. Added semantic/slice/identity/reordered-column tests and ASCII/Unicode combining/control character coverage. No driver/API changes.
• Exact current encoder microbenchmark: -Onone 2.95988s and -O 0.403835s per 100k rows / 35,862,200 SQL bytes. Prior optimized baseline was 3.202668s (~7.93× encoding-only improvement). Prior artifacts preserved in encoding-baseline/; new results in encoding-comparison/. This does not claim 8× end-to-end speed. Focused correctness regression is running (.build/transfer-optimized-regression.log); next build Release with ENABLE_TESTABILITY=YES and SWIFT_OPTIMIZATION_LEVEL=-O, run small live samples first, then Release 1M comparison followed by 10M. Benchmark CLI now defaults to Release and supports --configuration/--skip-cancellation. Keep source frozen during final performance matrix.
• The first 29-method regression exposed a bad reference expectation for combining-character strings: Foundation replacingOccurrences leaves ASCII quote/backslash untouched when followed by U+0301 (confirmed actual UTF-8 bytes), while PostgreSQL parses those ASCII bytes as syntax. The new UTF-8 encoding correctly doubles them. Split the ASCII reference and independent explicit combining expectation, and added combining characters to 1/17 of live fixture payloads. Final 29 methods all passed, 0 failures/skips (.build/transfer-optimized-regression-final.log).
• Source is now frozen. Release build running via XcodeBuildMCP, ENABLE_TESTABILITY=YES and SWIFT_OPTIMIZATION_LEVEL=-O (.build/transfer-release-build.log). Planned sequence: Release 1003-row four-direction smoke with combining payloads and no cancellation (too few rows to cancel deterministically), Release 1M four-direction validation/performance plus cancellation, then Release 10M four directions. No further Debug large matrices; no new production source changes until evidence warrants them.
• First Release dependency build still inherited coverage instrumentation from prior tests. Canceled that owned build, then a testing-only -enableCodeCoverage flag was rejected by xcodebuild build before compiling; removed it. Final build uses global ENABLE_CODE_COVERAGE=NO and CLANG_ENABLE_CODE_COVERAGE=NO along with -O and ENABLE_TESTABILITY=YES. Verified actual compiler commands have -O and no profile instrumentation. Final Release build succeeded in 317.9s (.build/transfer-release-build-final.log).
• Release host loaded/activated all 6 installed API 5 drivers successfully (.build/transfer-release-drivers.log; artifact 20261020-092034-236259). Started 1003-row four-direction live smoke (.build/transfer-release-smoke.log). Installer now accepts --configuration Release while preserving Debug as its default and records QCDevelopmentConfiguration, so final Dev can use the exact optimized configuration tested here. No commit/tag/public release or external driver changes.
• All four Release smoke cases passed 1003 rows with exact counts/ID sum and zero invalid values, including combining-character payloads (artifact 20261020-092101-7dde85). Their subsecond RSS sampling is not meaningful and must not be presented as real peak memory. Started Release 1M four-direction performance/validation (.build/transfer-release-million.log), skipping duplicate cancellation runs here; the final Release 10M run will include cancellation/overwrite preservation checks.
• All four Release 1M cases passed exact full-row validation (artifact 20261020-092307-6c548a): MySQL→MySQL 34.1072s/31.98MiB, PG→PG 32.7911s/33.40MiB, MySQL→PG 30.5176s/33.37MiB, PG→MySQL 36.7443s/35.78MiB. End-to-end rates ~27.2–32.8k rows/s; no claim that encoding's 8× applies to full transfer. Changes and compilation mode jointly improve on earlier Debug baselines.
• Followed user's request to avoid an oversized slow matrix: final 10M phase covers MySQL→MySQL and PG→PG plus append/overwrite cancellation; cross-engine correctness/performance is covered at 1M. Explicitly told user this scope. Started python3 scripts/benchmark-database-transfer.py --configuration Release --rows 10000000 --cases mysql-mysql postgresql-postgresql (session 85914), log .build/transfer-release-ten-million.log. Do not claim 10M cross-engine tests. Installed drivers remain unchanged API5 versions; final benchmark uses their exact installed bundles with the Release host.
• Final Release 10M artifact: 20261020-092837-0b7533. MySQL 5.7→8.0 completed and full-row validation passed: 376.345293s / 26,571.34 rows/s / 39.74MiB sampled peak RSS; max 100ms main-actor heartbeat gap 0.215s. Correct 10M rows, ID sum 50,000,005,000,000, zero invalid payload/decimal/raw/flag values. Earlier Debug 10M baseline was 1221.24s (~69.2% less elapsed with both code and compilation changes, not a pure optimization-only comparison). PostgreSQL 10M and 4 cancellation checks still running; no claim of completion yet.
• Final Release phase completed successfully (session 85914, exit 0): PG 17→17 10M took 321.825881s / 31,072.70 rows/s / 37.70MiB peak sampled RSS, max heartbeat gap 0.193s. Full target count/ID sum/all values verified, zero differences. MySQL and PG append cancellation verified zero target rows; overwrite cancellation verified the original sentinel row and every field were restored. Cancellation completion times: MySQL 0.725/0.827s, PG 0.109/0.107s. All three owned containers/volumes removed. Final scope is 4×1M + 2×10M + 4 cancellation cases, not 10M cross-engine.
• Documented methods, exact results, memory/UI limits, six-driver scope and reproduction commands in docs/DATABASE_TRANSFER.md. Benchmark default reduced to 1M; 10M now requires explicit --rows. Python syntax checks and git diff --check passed. No external driver modifications.
• Dev had 0 active transfers before installation. Installed/signed the tested Release app with normal quit via scripts/install-dev-app.py --configuration Release, backed up the prior app and launched /Applications/QueryCraftDev.app. Marker 20261020-094450-26feb753, QCDevelopmentConfiguration=Release, PID 37387 at verification. Installed QueryCraftFeature framework UUID F2C20B4D-C698-365E-AD59-869A26025B87 matches the tested Release product. Install log .build/transfer-release-install.log. No forced termination, desktop input simulation, commit, tag or public release. Task complete; only final user response remains.

Native bulk follow-up — 2026-10-20

• User requested the final complete optimization, not a list of further suggestions. Implementation includes native streaming reads, COPY OUT→IN for matching PostgreSQL layouts, bounded conversion streams, MySQL memory-only LOCAL callbacks, staged ordered Upsert, cancellation of active server waits and throttled UI progress.

• 41 focused methods passed before live validation. Release host and three affected drivers built/packaged successfully (v1).

• Live MySQL suite passed native append/upsert/overwrite/rebuild, FK/view/routine coverage, byte/Unicode/NULL preservation, wide-row chunking, SQL modes, duplicate/truncation warning rollback and active-query cancellation.

• Live PostgreSQL general feature suite passed. The new edge fixture initially lacked its catalog identity (test corrected). Rule-removal edge then exposed PostgreSQL's stale relhasrules optimization flag; changed runtime capability check to query actual pg_rewrite rows.

• Mixed old-Doris→new-MySQL loading reproduced a missing-symbol error in the old shared transport. New transport is now QueryCraftMariaDBTransportV2 (distinct module/install name, same source directory); both dependent drivers and metadata test imports were updated. This allows old V1 and new V2 bundles to coexist. Full mixed-load rerun remains pending.

• V2 unit and Release rebuild pipeline is running (.build/transfer-native-unit-v2.log, .build/transfer-native-release-builds-v2.log). No user app/installed drivers have been replaced yet; all Docker regression containers were removed.

• Final edge validation passed all four 203-row directions, full-SELECT binary validation, wide rows, rules/RLS/trigger fallback, raw COPY source errors, native warnings, active wait cancellation, and disabled-LOCAL streamed INSERT.

• Doris automatic structure tests then exposed decimalv3(30, 9); normalized DECIMALV2/V3 and DATEV2/DATETIMEV2 names. Read-only/data-only Doris checks had already passed.

• Added safe staging availability fallback: a SAVEPOINT protects temporary-table preparation so DML-only accounts can still Upsert after a staging permission failure. Added a unit case and real restricted-user tests for MySQL/PostgreSQL.

• 45 focused methods passed (.build/transfer-native-unit-permissions.log). Latest Release host rebuild runs in exec session 78267; no installed app/driver replacements yet. The earlier Doris fixture was removed by validation cleanup, and the final validator now creates and owns a fresh name-isolated fixture each run.

Native bulk transfer completed — 2026-10-20T11:49:11

• Implemented optional API 5 COPY/controlled in-memory LOAD DATA writers and bounded source streaming; PostgreSQL compatible raw COPY relay; transactional indexed Upsert staging and SAVEPOINT-protected permission fallback. Preserved exact values, warning failures, ordering and rollback. No filesystem LOCAL access.
• Final 45 regression methods passed. Live four-direction edges, DML-only users, disabled LOCAL fallback, six-driver loading and three real Doris/mixed-version combinations passed. Shared MariaDB transport uses V2 module/install name so old/new Doris and MySQL can coexist.
• Release million append/Upsert: 8 cases, all full-row values correct. Release ten-million append: MySQL→MySQL 102.057 s / 36.62 MiB; PG→PG 45.746 s / 26.34 MiB; MySQL→PG 48.941 s / 40.05 MiB; PG→MySQL 121.809 s / 40.62 MiB. RSS is one-second engine-harness sampling, not full app memory. Native path proven by progress messages.
• Six real cancellation cases passed with zero committed rows; overwrite restored sentinel, staged Upsert restored all 1M original rows. MySQL cancellation took 0.722/0.724/4.280 s; PG 0.109/0.109/0.107 s. All owned Docker containers/volumes removed.
• Artifact roots: .build/validation/transfer-performance/20261020-113229-5a65b4 and 20261020-113616-289698; include results, logs/RSS, environment and host/driver fingerprints. Full results and limits: docs/DATABASE_TRANSFER.md. Doris real small samples only, MariaDB server untested; Redis/ES/Kafka outside relational transfer. No visible UI/physical-input verification.
• Installed and launched signed Release Dev 20261020-114655-8a63b4d1, feature UUID 3CDC441A-DED7-3F94-8738-401A8C71EC0C, MySQL/PG/Doris local 1.0.6 build 8 (API5, min app0.1.10). Installed full-bundle hashes and six driver binary hashes match benchmark artifacts; installed-host six-driver load passed; actual Dev PID68023 loaded six installed drivers and V2 framework. Normal quit only; old app/driver backups retained. No commit/tag/public release.

0.1.11 release preparation

All main features merged, Tools commands unified, app 0.1.11(21), MCP0.1.1(2), MySQL/Doris1.0.7(8), PG1.0.6(8). Other drivers and Skill unchanged. Integrated 227 Release regressions, 15 scope tests, Release build, all12 online driver archive checks, standalone Release runtime/six/mixed loading and isolated Debug lifecycle passed. Release HTTPS policy correctly rejects loopback component fixtures; Debug is required for those fixtures. A duplicate local packaging process was resolved and resulting Debug signature rechecked. CI now gates both native architectures on packaged runtime/driver checks.

iOS / iPad proof — 2026-10-22

• Read repository instructions, planning, SwiftUI, concurrency and XcodeBuildMCP skills. Created/attached managed worktree. No physical desktop input.

• xcodebuildmcp and xcodegen are available. Investigating installed iOS runtimes and portable transport source closure.

• iOS 26.3 runtime installed by user; iPad Pro 13-inch and iPhone 17 simulators available. Native x86_64-simulator OpenSSL 3.6.3 and MariaDB 3.4.9 built from checksum-pinned sources. arm64-device build in progress.

• Added independent XcodeGen mobile core/transport/application targets reusing original source files; desktop types remain in Feature. Only session-capability declarations were split from portable stream implementations.

• Implemented connection form, SQL editor, results and stable status/footer; endpoint changes and stale responses isolated. Added deterministic request-state tests.

• Build-script first Configure attempt rejected split -target arguments; fixed to --target= with immediate configure failure propagation.

• Final iPad suite: 14 passed / 0 failed / 0 skipped. Device build passed. Captured actual Chinese SQL results in .build/ios-validation/ipad-query.jpg. iPhone compact launch verification and Mac build regression in progress.

• Mac dependency resolution initially fetched unnecessary submodules/newer compatible versions; stopped only the isolated build and copied the existing local package cache into the worktree, restoring pinned versions. No primary source or installed Dev changed.

• Both native simulator architecture slices and arm64 device libraries passed checksum-cache revalidation. Added reproducible isolated Docker/MySQL integration script with cleanup trap; exercising it on iPhone as a second form factor.

• Rebuilt final arm64 device target without warnings. Shared source split compiles in the desktop Feature target; full desktop link is in progress.

• iPhone first test launch failed before XCTest bootstrap (signal kill during initial boot / concurrent launch activity; root cause not confirmed). Serialized the launch sequence and retried after device boot; first attempt is not recorded as an app test pass. Provisioning-script cleanup removed its random fixture.

• iPhone retry and fixture script passed all 14 tests, with automatic random container/volume cleanup. Full Mac Debug build passed. All six installed API 5 drivers activated and created session factories against the newly built Feature in an isolated CLI process.

• MySQL and Doris desktop targets independently compiled successfully from the worktree. MySQL local ad-hoc bundle/archive passed package signature checks; Doris packaging in progress. Installed Dev remains untouched.

• Final verification complete: iPad and iPhone each 14/14, final unsigned arm64-device build, full Mac Debug build, six installed-driver factory compatibility, fresh MySQL/Doris builds/packages/signatures and mixed loading. Fresh MySQL package also connected with TLS, queried 2 owned rows and preserved alias origin.

• Final artifacts/limits recorded in QueryCraftMobile/README.md and VALIDATION.md. Original ephemeral MySQL fixture and anonymous volume removed. No physical desktop input, installed Dev replacement, user database writes, commit, tag or release.

MySQL mobile first usable version — 2026-10-22

• User authorized implementation and then prioritized MySQL only. Reused the existing attached worktree and read current source/plan/instructions. Earlier prototype changes are preserved.

• Applying the existing planning, SwiftUI expert, Swift concurrency and XcodeBuildMCP workflows. No new agent delegation or desktop input.

• Deferred PostgreSQL/Doris/Redis/Elasticsearch adapters before any edits to them; Kafka remains out of scope.

• Added additive bounded native query previews with raw binary/null cell metadata and socket shutdown before freeing truncated unbuffered results; existing desktop query API/layout unchanged. First simulator compile passed.

• Added MySQL metadata/page service, database/object/table request state, read-only single-statement policy using actual sql_mode, native Keychain abstraction and protected atomic profile/history repository.

MySQL mobile implementation and UI follow-up (2026-10-22)

• Implemented saved connections / optional device-only Keychain credentials, database/table search, schema metadata, server-side paging/sorting, conservative read-only SQL preview, history and cell/row inspection. Native preview closes truncated unbuffered queries instead of draining all remaining rows.

• Shared desktop sampling and long-cell preview without changing Mac behavior; native 2D reusable grid with generation-based updates and preserved scroll/selection/widths.

• Each simulator passed 32 tests before final visual refinement. Actual Keychain test exposed unsigned simulator entitlement failure; enabled ad-hoc signing for simulator and verified read/write/update/delete.

• User requested more TablePro UI reference, then objected to the rough appearance. Read its mobile UI sources; independently implemented compact record summaries/detail, connection/metadata hierarchy, native grouped schema list, readable table metrics and OS26 Liquid Glass action controls with older-OS fallback. Final tests/rendering and driver package regression remain in progress.

• Final checkpoint: each simulator passed 32/32 (0 skipped) after UI refinements; SQL/data/schema snapshots inspected. Final ARM64 device build passed. Mac and fresh MySQL/Doris builds/packages, six old-driver compatibility, mixed new/old loading and new MySQL real TLS/metadata query passed. No installed app/driver replacement. All owned test containers and anonymous volumes cleaned; diff whitespace and script syntax checks passed.

2026-10-22 — Clean connection home and native navigation tabs

• Separated connection home (saved endpoint cards/add entry/first connection state) from the database workspace; password entry gates opening, reopening the current connection retains its query draft/result and loaded catalog.
• Replaced catalog SQL row with native TabView Tables / Query / Views. iPhone uses the system bottom Liquid Glass tabs; iPad uses system top tabs with object split navigation. Database selection is available at the top of each context.
• MobileMySQLService filters BASE TABLE / VIEW before LIMIT; kind changes cancel and isolate catalog/table state and guard late responses. Returning between query and the same object kind preserves cached object data; query drafts/results remain in the workspace. Top refresh follows the active destination.
• Final simulator runs: tabs-iphone-tests.json 37 passed/0 failed/0 skipped (22.808s); tabs-ipad-tests.json 37 passed/0 failed/0 skipped (19.454s). Real MySQL fixture validates separate table/view catalogs and reading a VIEW. ARM64 unsigned device target builds (tabs-device-build.json), no warnings. Native screenshots inspected for iPhone catalog/views/query/table/structure and iPad catalog/views/query/table.
• Home runs before tab steering also passed 35 tests each; final home screenshots inspected. No desktop input simulation used. Real touch/IME/rotation and signed-device execution remain unverified. Owned test containers and volumes cleaned. No commit/tag/release.

2026-10-22 — Physical iPhone installation

• User connected and enabled Developer Mode on an actual iPhone. Used XcodeBuildMCP CLI; no desktop mouse/keyboard simulation.

• Read-only verified the existing Apple Development certificate team, automatic development profile, app/team entitlements, connected-device inclusion and strict signature verification. Personal signing settings were command overrides only.

• Generic + concrete destination build returned a generic provisioning failure but produced a complete signed concrete-device app. Independently verified that product before installing. XcodeBuildMCP device install succeeded (.build/iphone-device-install.json).

• Launch remains denied by iOS security pending manual developer-profile trust (.build/iphone-device-launch.json). Requested user confirmation via phone Settings. Do not claim actual device launch or interaction testing complete yet. Profile expiry 2026-10-29. No commit/tag/release.

• Follow-up: user confirmed trusting the developer App. Device launch succeeded with process ID 665 (.build/iphone-device-launch-trusted.json). Installation and launch are complete. Actual mobile database interaction remains for user testing and is not counted as verified by a successful launch command alone.

2026-10-22 — Shared desktop search, compact data detail and phone landscape

• Persisted the requested mobile/desktop functional baseline and shared-logic requirement in both primary and managed-worktree AGENTS.md. No other primary-checkout changes.
• Located the actual Mac object-list path and extracted its existing matching/sort/tie handling into CompletionLabelMatcher.matching. Both platforms compile that same source and call that method. Mobile catalogs load complete bounded metadata and filter locally; ad/amu/case, Unicode, literal wildcard symbols, stable ranking, >200 objects, local typing without requests and refresh retention are covered.
• Native grid is the default on both mobile sizes. Replaced floating controls with a system top menu, retained query history there, moved Data/Structure into the title menu, and kept one status/pagination footer. Hide main tabs only in compact detail layouts; regular iPad keeps top tabs to preserve query/view navigation.
• Added standard iPhone portrait/left/right orientation declaration; iPad retains four orientations. Actual right-landscape phone render and native viewport/header/width/selection assertions passed. The iPad scene request left its window portrait; did not count that image as physical landscape validation.
• Final evidence: compact-detail-iphone-tests.json 43 passed/0 failed/0 skipped (16.880s); compact-detail-final-ipad-tests.json 43 passed/0 failed/0 skipped (25.464s). Wide fixture has 78 columns/120 rows. compact-detail-final-device-update.json signed build/install/launch succeeded (18.503s), existing profiles retained, no debug fixture environment supplied to phone.
• Mac full build succeeded (clean-grid-mac-build.json, 477.245s). QueryCraftFeature scheme test attempt could not execute because no test action is configured; it is explicitly not a pass. An isolated executable linked the actual compiled framework and invoked WorkspaceDatabaseItems.objects, validating table/view/ES resource kinds, requested admin_user fuzzy cases, rank/ties/empty order, and original bo regression (ios-validation/mac-search-check.txt).
• Existing mixed six driver bundles all activated and created sessions against the newest worktree Feature framework; shared sampling/cell preview assertions also passed (ios-validation/compact-detail-six-driver-load.txt). Driver API and protocol conversions are unchanged; no driver repackaging required for this private shared-search extraction. Other drivers did not receive mobile support.
• Final portrait/landscape/structure/query/search screenshots inspected. No desktop input simulation used. Owned fixtures and volumes removed; whitespace validation passed. User real-device gesture/IME/sensor rotation checks remain manual. No commit/tag/release.

2026-10-22 — Fullscreen data grid and strict single-axis panning

• Table and SQL result menus now offer Fullscreen Grid. A native full-screen cover contains the existing renderer and current typed page/result only; no navigation, sidebar, editor or status/pagination/footer. Exit auto-hides after 2 seconds, tap reveals it, Escape and accessibility escape dismiss; initially enabled VoiceOver keeps it visible.
• Added non-observable viewport state and owner tokens. Full-screen and normal grids preserve offset/widths/selection; inactive background callbacks cannot overwrite active full-screen state. New object contexts reject old viewport data. Full-screen does not load other pages or query again.
• Enabled native directional lock and added strict dominant-axis constraints for offsets and deceleration targets. Tiny ambiguous movement waits; each gesture can choose a new axis, including after cancellation. Same-object refresh preserves an ongoing lock. Column resizing recognizers are enabled only for actual headers and accept horizontal intent, so data cells no longer steal swipes.
• Final results: fullscreen-final-iphone-tests.json 49 passed/0 failed/0 skipped (18.054s); fullscreen-final-ipad-tests.json 49 passed/0 failed/0 skipped (16.014s). Added four Swift Testing axis regressions and two native UIKit viewport/delegate regressions, including same-object refresh during a gesture and full-screen restoration/background-write fencing. Real isolated MySQL and previous regressions remain covered.
• Inspected table/query fullscreen on phone and iPad, phone right-landscape fullscreen, and settled iPad render with no persistent exit control. No desktop input simulation or real-device synthetic gestures used. Manual physical diagonal-drag/deceleration/rotation acceptance remains for the user.
• Signed ARM64 build's original log says BUILD SUCCEEDED and codesign --verify --deep --strict passed. Combined build-and-run tool reported failure because launch was denied by a locked phone. Separate install succeeded (fullscreen-device-install.json). User told to unlock and open the updated app; did not claim startup success for this latest update. Existing connection data was not cleared.
• Mobile-only changes apply to MySQL table and SQL grids, normal and fullscreen. Mac layout/wheel behavior, public models, external driver protocols and the other five drivers are unchanged. All owned fixture containers/volumes removed; script syntax and whitespace checks passed. No commit/tag/release.

2026-10-22 — Fullscreen safe area and keyboard follow-up

• Removed fullscreen grid's all-edge safe-area override; only its background now extends under system/hardware regions. Catalog search has explicit focus binding/reset on navigation, results/editor clear focus before fullscreen, editor exposes a keyboard dismissal action, and grid dragging dismisses input.
• Added presentation-window-only responder guard: dismiss on entry/size changes and late keyboard show, release observers/pending callbacks on removal. No process-global resignation, no data/search clearing, no desktop/driver changes.
• First test build failed on an explicit-self requirement in the new recursive UIKit test helper; corrected. CLI returned a structured failure without nonzero exit, so fixture script now checks JSON success before rendering to avoid stale-binary screenshots. This failed attempt is not a test pass.
• Final iPhone and iPad suites each passed 51/0/0 (18.424s / 22.396s), including two new native UI regressions for safe layout/pinned header during scroll and asymmetric resized proposals, and retained responder dismissal/release. New portrait/landscape phone and table/query iPad screenshots inspected; no app bars or keyboard in fullscreen, header/cells stay clear of notch region.
• Signed device build/install/launch succeeded (safearea-keyboard-device-update.json, 21.051s, PID 1112), no fixture launch environment passed and no saved connections cleared. User's physical third-party IME/rotation check remains manual; do not equate responder tests with that acceptance.
• Final whitespace and fixture-script syntax checks passed. Owned MySQL containers/volumes removed. Mobile-only fix covers table and SQL grids, catalog search/results search/editor focus; no desktop/shared models/drivers changed. No commit/tag/release and no desktop input simulation.

2026-10-22 — User simulator experience

• User asked to experience in simulator. Opened Simulator.app and launched the already verified latest iPhone 17 Pro app in normal table mode (no automatic fullscreen), connected to an owned loopback-only read-only MySQL experience fixture with 78-column admin_user and 120 rows. Runtime launch succeeded (PID 42350), existing user databases untouched.
• Deliberately keep this experience fixture alive for user interaction; owned container name and loopback port are recorded only in ignored .build/simulator-experience-{container,port}.txt. Remove this named container/anonymous volume when the user finishes the experience or no longer needs it. Do not remove before the user can refresh/page/query.
• User also requested iPad experience: launched the latest app on booted iPad Pro 13-inch (M5), same live owned fixture, normal table mode (no auto-fullscreen). Launch succeeded (PID 44301); Simulator.app opened. Phone experience and read-only fixture remain running for the user.

2026-10-22 — Shared grid find parity

• Split desktop source enum into its own adapter and introduced a Sendable source protocol; mobile compiles the existing worker/controller/operator/match source files directly. Desktop convenience overloads retain existing inferred source call sites; matching, debounce, cancellation, generation guards, cap and next/previous state logic are unchanged.
• Mobile now selects/reveals matching cells, preserving all original rows and raw typed values. Added column/operator/case options and wraparound navigation in one opt-in 44-point bar, keyboard shortcuts, weak UIKit selection handler, resize visibility and fullscreen owner checks.
• Initial phone suite passed 59/0/0. Render review caught landscape keyboard crowding, including SQL editor taking result space and a simulator keyboard onboarding overlay. Added compact input layout, explicit hide-keyboard/finish-find control, and collapse SQL editor while result find is shown. Debug rendering now shows after-dismissed keyboard state without synthesizing desktop input; no physical keyboard behavior falsely marked verified.
• Mac full Debug build passed (grid-find-mac-build.json, 160.001s). Actual compiled desktop adapter/operator/controller regression passed, including a disk-backed 50,002-row query source with a match in its final row; all six existing mixed driver bundles activated and created sessions against the new Feature, with live owned MySQL TLS/alias-origin query passing.
• Added a query-editor collapse UI regression. Two attempts incorrectly asserted that SwiftUI must destroy a TextEditor UIKit object; the runtime retained it while offscreen. A separate actual app render confirmed editor space is reclaimed and matching cell visible. Changed the regression to assert grid geometry and draft/state restoration, rather than object deletion; earlier failed runs are not passes.
• Follow-up corrected the UI-test diagnosis: a third geometry check showed the test's disconnected legacy UIWindow was not driving SwiftUI observation updates. Attaching the test host to the active UIWindowScene exercises the real update cycle; the targeted editor-collapse/draft-restoration regression now passes. No product workaround or removal of the regression.
• Final focus-scoped resize guard prevents an old find selection from moving the manually restored viewport after fullscreen/rotation. Regression covers both active keyboard resize visibility and post-input resize preserving a manually scrolled offset.
• Final iPhone and iPad each passed 60/0/0 (20.860s / 24.637s), evidence grid-find-final-phone-tests.json / grid-find-final-ipad-tests.json. Inspected table-find, query-find and fullscreen native renders on both; preserved snapshots at ios-validation/grid-find-{phone,ipad}-{find,query-find,fullscreen}.jpg. Screenshots show landscape UI in portrait pixel axes; no physical input was synthesized.
• Latest signed device build/install/launch succeeded (grid-find-final-device-update.json, 18.586s, PID 1194); strict signature verification passed. Earlier locked launch is retained as failure evidence, superseded by this successful update.
• Restored both simulators to normal table mode on the user's retained read-only experience fixture, with no automatic find/fullscreen flags. Test fixtures are removed by script cleanup. Shared Mac behavior remains intact; mobile find scope remains loaded page/result, other mobile drivers deferred. Documentation and script syntax/whitespace checks updated; no commit/tag/release.

2026-10-22 — Shared timestamp display parity

• Traced actual Mac formatter, header action, table scope and query store scope. Mobile compiles the same formatter; conditional Core import and equivalent origin lookup leave desktop behavior intact. Transport workspaceColumns now reaches mobile grid metadata instead of being discarded.
• Native header context menu and accessibility actions select automatic/raw/seconds/milliseconds. Only the affected column is resized; non-observable format state is shared through fullscreen viewport ownership. Background updates cannot reset active fullscreen formats. SQL resultRevision is a separate display scope, so each successful new result resets modes like the Mac query store.
• Added five Swift Testing cases for shared formatting/type/origin/bounds/time zones/state/raw search and real MySQL table/aliases/cross-database joins/expressions/empty metadata, plus two native XCTest UI regressions for menu actions/fullscreen/refresh/new-result scope/viewport.
• Early test compile attempts referenced a private cell type and omitted a required nullDisplayText argument; corrected without changing product access. An initial native regression directly dequeued datasource cells and triggered UIKit's consistency check; switched to actual visible collection-view cells. These attempts remain failures, not successful validation.
• Phone/iPad initial 67-case suites passed. Screenshot review caught pre-existing compact-landscape SQL editor consuming all result height. Reduced only compact-height editor to 60 pt and removed its duplicate result heading; regular layout remains 130 pt. Existing hosting regression now also asserts usable result height.
• Final phone/iPad each 67/0/0 (84.745s / 32.861s) after this layout fix; final reports timestamp-final-phone-tests.json / timestamp-final-ipad-tests.json. Mac full Debug build passed (598.306s), native compiled table/query formatter/raw copy/export/lifecycle regression passed, and six driver loading/session factories plus owned MySQL TLS query passed. Device signed build/install/launch succeeded (79.830s, PID 1227), strict signature verification passed.
• Inspected final phone/iPad timestamp table, alias-query and fullscreen screenshots; both render formatted dates, full raw 18-digit IDs and unconverted varchar dates correctly, with no fullscreen overlays after timeout. Final snapshots retained at ios-validation/timestamp{,-query,-fullscreen}-{phone,ipad}.jpg. No physical input simulation was used.
• Extended only the owned live experience fixture with epoch_values, then restored phone/iPad to that normal table page (no auto-find/fullscreen) with successful launch reports timestamp-{phone,ipad}-experience.json. Testing fixtures/volumes removed; experience fixture intentionally remains. Docs, script syntax and whitespace checks passed. No commit/tag/release.

2026-10-22 — iPad sidebar top alignment

• User compared iPad Contacts with our screenshot: root TabView reserved a full-width tab region above each nested NavigationSplitView, lowering the entire sidebar. In regular width, moved NavigationSplitView to the workspace root and confined native tabs to its detail. Compact width retains the existing tab/navigation hierarchy.
• Regular sidebar remains shared across table/query/view tabs, using the browser's actual catalog kind; object selections and all refresh/navigation behavior keep the same application model. No negative padding or safe-area bypass. Added a native hierarchy/top-toolbar regression plus tab identity/draft preservation checks.
• User then asked to remove the tab row as it still consumed too much space. Final regular-width layout removes TabView altogether, using the existing sidebar navigation-title menu for Tables/Query/Views and a single native detail navigation container. Compact width keeps the existing system tabs. This also avoids nested detail navigation containers swallowing toolbars.
• Regular-width hierarchy/no-tab regression remains meaningful on both simulator families; restricted top-safe-area geometry to actual regular height after the forced regular-width phone test received native compact-height status-bar insets (24 pt). Intermediate phone geometry failure was not treated as a product pass.
• Final title-menu suites passed 68/0/0 on iPad (42.433s) and iPhone (25.875s), sidebar-menu-{ipad,phone}-tests.json. Inspected regular table/query/views showing top-aligned sidebar, no tab row, native table title/data-structure menu and refresh/results tools restored; compact phone retains native tab/back routes.
• Signed device build/install/launch passed (sidebar-menu-device-update.json, 22.202s, PID 1257). Both simulators restored to normal epoch_values experience using the retained fixture; iPad requests right-landscape scene geometry; raw capture pixel orientation is not physical rotation acceptance. Snapshots sidebar-menu-{ipad,phone}-*.jpg retained. No real desktop input simulation was used, no Mac/driver/data changes and no commit/tag/release.

2026-10-22 — Shared MySQL server filtering

• Extracted MySQLWorkspaceSession's actual SQL/binding predicate builder to a driver source file compiled directly into mobile; the desktop keeps bound values, mobile's bounded preview uses UTF-8 hex string literals through the same predicate construction. No transport query/preview ABI or Driver API change. Mac default-condition, enum/bool choices and column-selection reset rules are shared through WorkspaceSQLDataFilterColumn; mobile directly compiles the existing filter models/operators/validation/titles.

• Native optional filter sheet uses stable condition UUIDs, same type/operator rules, enabled conditions, all/any logic, enum/bool choices, add/remove/clear, and independent cancelable draft. Menu-only entry and existing status footer; applied page stores its filter separately from requested filter. Server WHERE precedes sort/LIMIT/OFFSET. Failure retains old grid and committed page/filter; retry starts at page zero for pending new filter; stale replies/object changes/cancellation fenced. Fresh schema is exposed for repairing invalid filters while old grid remains.

• First simulator build hit Swift 6.2.4 IRGen crash converting an actor-isolated method reference to a Binding setter; replaced it with an explicit setter closure and main-actor apply callback. Retry build succeeded. Initial phone/iPad each passed 79 tests; screenshot review moved modal explanation to bottom and declared zh-Hans development language to keep shared labels consistent with the Chinese mobile UI. Updated 79-case suites passed again (phone 28.100s / iPad 34.844s).

• Traced Mac viewport application too: same columns/offset/limit/sort/filter preserves refresh position, changed context clears selection and scrolls to row zero. Added committed sort provenance and a typed mobile page context, carrying it through ordinary/record/fullscreen views. Keep horizontal position/widths/timestamp modes while a new logical page resets its row; restore only matching page/metadata. Also avoid redundant same-filter reloads. Added native regression; final expected count 80.

• MySQL driver rebuilt successfully (739.405s), packaged/signature checked in isolated server-filter-driver-mixed output, archive hash prefix 4903c50bc59f316f. Native compiled Mac default helpers and newly packaged MySQL live TLS filter/paging/sort/count/LIKE/null/empty/injection-like string checks passed, all six packaged drivers loaded and created sessions. The harness uses the actual Dev registry/factory API against the isolated new bundle; installed Mac/Dev app and driver library were not replaced.

• Final iPhone suite passed 80/0/0 (27.125s). User requested fixed labels left and values/inputs right; changed both first and second fields to native LabeledContent. Final iPad/phone each passed 80/0/0 after this change (26.763s / 18.353s); actual native filter forms and rows 110-115 renders inspected. iPad raw image axes remain inverted; no physical rotation acceptance claimed.

• Mac full build passed (555.836s); native compiled defaults + six Dev registry driver factories + rebuilt MySQL TLS/live filtering rerun passed after full main build. Latest signed device build/install/launch passed (filter-value-alignment-device-update.json,18.641s,PID1392); strict signature passed. Both simulators restored to retained live filtered fixture without automatic editor/find/fullscreen. Test fixtures cleaned, experience fixture retained; no installed Mac/Dev library replaced. Docs/limits updated; no commit/tag/release or physical desktop input simulation.

2026-10-22 - Explicit filter keyboard dismissal

• User could not discover how to dismiss the filter keyboard. Retained interactive Form swipe dismissal and added a native keyboard-toolbar Done button, with a distinct focus identifier for each condition/value bound. Return also dismisses focus, without applying or discarding the draft. Applies to both iPhone and iPad; no shared database/driver/Mac behavior changed.
• Simulator build/run passed (filter-keyboard-simulator-update.json,15.663s), latest app installed/launched on both simulator families with live fixture restored. Device build log succeeded and strict signature passed; combined action failed only at locked launch. Separate device install succeeded (filter-keyboard-device-install.json); user can unlock and open app. No claim of actual touch/third-party keyboard validation. Whitespace check passed; no test suite repeated for this small focus/UI change.

2026-10-22 - Prefer the keyboard Return action

• User rejected the extra Done accessory and clarified that the keyboard bottom-right action should dismiss input. Removed accessory toolbar, retained per-field submitLabel(.done) / onSubmit focus clearing. No automatic apply or draft loss. Native scroll dismissal is immediate; no custom drag recognizer or extra sheet grabber added after clarification. Applies to both mobile sizes; database/Mac/drivers unchanged.
• Final native simulator build/run passed; latest signed device build/install/launch succeeded (filter-return-keyboard-device-update.json,11.495s,PID1429), strict signature passed. Prior intermediate locked launch remains recorded as failed. No test-suite rerun for this small UI change; actual soft-keyboard/IME confirmation remains with user. Both simulator apps updated with same final source.

2026-10-22 - Shared loaded-data export

• Added shared writer/request/options/preferences/dialect sources to mobile, extracted unchanged copy range/column/snapshot types from AppKit-only encoder. All four formats use original writer, default JSON objects and no SQL transaction wrapper. Found/fixed shared JSON alias suffix collision while retaining ordinary duplicate naming.
• Implemented immutable committed raw snapshot, native options/save picker, safe generated filename, cancel/generation fencing and per-attempt directories. Avoid old cancelled writers deleting a newer file. Explicit loaded-page/result scope, SQL target naming parts, all format options behind disclosure; no new toolbar row. No remote writes/driver API change.
• Initial mobile build passed. First new test build rejected a non-Sendable predicate crossing actor boundaries; made the polling predicate explicitly Sendable. Earlier attempt is not a pass. Shared/live MySQL format, binary/NULL/raw-value, atomic failure and noncooperative cancellation regressions added.
• Phone/iPad each passed 89/0/0 (32.236s / 27.608s), including a native Files-picker source lifetime/cancel/retry regression. Inspected table/query export forms on both. Final footer explains binary summaries rather than promising raw bytes; final simulator/signed device builds passed (16.785s / 12.737s), strict signature passed. CoreDevice installation failed due disconnected/unavailable phone, independently confirmed; not counted as installed.
• Mac full Debug build passed (194.816s); actual compiled four-format/default/dialect regression passed, independent Python ZIP/XML inspection confirms raw long integer/text/header behavior. All six driver registry/factories and owned MySQL TLS checks passed against updated Feature. Host-only changes do not require repackaging external drivers; installed Mac/Dev/library untouched. README/parity/validation and limitations updated; no commit/tag/release or desktop input simulation.
• Final footer render caught the binary disclosure being omitted when Section footer had multiple root children; wrapped in VStack. Phone rerun passed 89/0/0 (27.111s), binary-summary text visibly rendered. Final signed ARM64 build passed (8.016s), strict signature passed. Same binary installed in iPad; final experience restored after fixture tests. No owned test containers remain; experience fixture intentionally retained. Final whitespace/script checks passed.

2026-10-22 - Shared SQL editor

• Compiling Mac structural parser/completion/ranker/query resources and default palette directly into mobile; native UITextView adapter preserves UTF-16 selection, IME text, undo and scroll. Contextual candidates use a recyclable native keyboard accessory, touch or Return/Tab, Ctrl-Space and Cmd-Return; no permanent suggestion toolbar.
• Added iOS declaration to portable language package. First manifest attempt used v18 unavailable under tools 5.9; changed to string platform version. Its copied root Resources directory was rejected by iOS bundle signing; copy SQL/JSON query subdirectories instead, using existing resource lookup fallback. Missing Core import in parse snapshot corrected; selector named next collided with UIResponder getter, renamed.
• Extracted Mac MySQL catalog queries/conversion unchanged into driver-owned shared helper; mobile reads the same real schemas/fields/comments/ordinals lazily. Metadata generation/cancellation/cache/error handling separate from query results. Original Mac completion tests and parser tests now compile conditionally on mobile too. Native editor and real metadata/stale/failure tests added.
• First full mobile suite compile required portable fixture-bundle lookup for the original Mac parser tests; added the same 88 KiB fixtures to the native test bundle. Phone/iPad each then passed 204/0/0 (46.088s/46.766s), including native IME/undo/UTF16/appearance and real MySQL lazy metadata. Actual SQL rendering uses shared Mac palette. Mac full build passed (203.631s), ARM64 signed build passed (60.778s).
• Further parity check retained Mac's active-relation column eviction, coalesced in-flight same-reference loads and fenced cancelled inactive relations. Long editor text applies colors only to visible composed-character ranges and refreshes on scroll, while parsing stays on its actor. Added cancellation/coalescing regression and final native completion probes; final expected suite count 205.
• New long-draft viewport regression initially compared a manually created UITextView using its default font with the production configured editor; line-height/content-size estimates shifted after per-range coloring. Factored the existing production factory and made native tests instantiate that exact view, retaining the viewport assertion. The failed iPad 206-case attempt is not a pass; targeted verification and final suites follow.
• New MySQL driver build passed (406.998s), isolated ad-hoc package signature passed (hash prefix 9f1e962816afe904). Actual compiled Mac parser/highlight/completion/IME application guards passed for MySQL/PG/Doris. Six packaged driver factories and newly packaged MySQL TLS catalog/view/78-field ordinal metadata and prior filter/paging checks passed through the real Dev registry/session API; installed Mac/Dev library untouched.
• Production-view native targeted suite passed 5/0/0 (22.778s). Full iPad viewport/cache suite then passed 206/0/0 (33.756s). Additional current-source dismissal guard keeps Escape-dismissed candidates closed across late metadata, and composing IME hides candidate contents without reloading its input view or intercepting IME keys; Cmd-Return commits composition first. Final phone acceptance passed 206/0/0 (31.931s), final signed ARM64 build passed (9.579s); iPad acceptance follows. No actual desktop input injection used.
• Final same-source phone/iPad acceptance each passed 206/0/0 (31.931s / 32.248s), including active-IME accessory protection, Escape/late response and native configured-font long viewport. Signature passed; final physical device list remains disconnected/unavailable, so latest SQL editor not installed on phone. Both simulator live query experiences restored without automatic keyboard/completion; retained fixture intentional. Source/planning/parity/validation docs updated, whitespace/script syntax passed. No commit/tag/release or actual desktop input injection.

2026-10-22 - Shared whole-table export

• Traced the real Mac grid allRowsProvider and separate direct-table target. Extracted the unchanged paging/schema/prepare/finish behavior into WorkspacePagedDataExportRowSource; Mac adapter retains its independent session, 5,000-row batches, PostgreSQL identity priming and complete-value validation. Mobile directly uses that common actor with its existing 200-row/8-MiB read-only table service, captured committed filter/sort/object/configuration, and raw cells.

• Added range picker to the existing export Form, default loaded-page scope retained, whole matched rows restarts at zero; query preview exports remain bounded and never rerun arbitrary SQL. Empty results and views supported. The shared writer prefetches one batch, not all pages. Changed native Files saving to URL-based UIDocumentPicker instead of loading a whole output into FileWrapper; native save/cancel lifetime test passes.

• First phone/iPad suites each 212/0/0 (39.293s/33.380s), including real 455-row multi-page/table/view/filter/sort/four-format/empty/oversized-row/cancel/schema checks. Added explicit failed second-batch retry and empty-scope model checks; final phone 213/0/0 (33.517s). Final iPad, Mac regression and simulator restoration follow. Signed ARM64 build passed (21.417s), strict signature passed; physical device disconnected/unavailable, no install attempted.

• Final iPad also 213/0/0 (33.516s). Mac full build passed (391.663s); actual compiled Feature checks verified all three SQL dialect batch boundaries/filter/sort/four formats, direct-table generated/identity preparation, binary/NULL rules, six driver factories and actual Dev registry MySQL TLS source/export. Independent CSV and XLSX ZIP/XML reads confirmed all 1,050 rows and order. No driver source/protocol changes or new driver package needed for this increment. Both simulators restored to retained live admin_user fixture; test containers/volumes cleaned. Final docs/checks complete, physical disconnected device not updated. No commit/tag/release or actual desktop input injection.

2026-10-22 - Shared MySQL data editing

• Portable column-array request preparation keeps existing Mac signatures as wrappers; same original-PK/null/default/generated/binary rules. MySQL driver and mobile share the transactional delete/update-group/insert runner and real bound statements; ROLLBACK uses unbound transport cleanup so cancelled tasks still roll back.

• Added per-object staged edits/inserts/deletes, raw-key mapping, pending cell/delete highlighting, existing-footer pending action and native compact cell/insert/preview forms. Safety lock starts on; explicit unlock/submit follows actual Mac behavior. Unknown COMMIT status blocks repeat submission pending server verification. Nontransactional engines are explicitly unsupported on mobile to preserve rollback semantics; query editor remains read-only.

• First build failed because newly created source files had not yet been regenerated into XcodeGen project; regeneration fixed it. Simulator build succeeded (19.150s). First iPhone full suite 220/0/0 (48.550s), including actual writable isolated MySQL tests: bound unusual/Unicode/NUL values, exact long/composite PKs and grouping, NULL/default/generated/binary/no-PK guards, insert/default/delete, whole-transaction rollback and cancellation, uncertain commit/schema gates. Continued lifecycle/native entry and staged-insert correction review before final acceptance.

• iPad intermediate suite 222/0/0 (52.698s). Added native grid/keyboard regressions and actual SELECT-only account write denial; the native vertical SwiftUI TextField Done test failed (raw log/result bundle retained; CLI failed before writing its summary), exposing a real keyboard bug. Replaced only data-value input with a native UITextView adapter that preserves multiline/IME/undo and resigns itself on Done; targeted native suite passed 2/0/0 (28.814s).

• Final acceptance phone/iPad each 224/0/0 (28.027s /31.734s). Signed ARM64 build passed (17.948s), strict signature passed, physical phone remains disconnected. Mac main build passed (148.651s); MySQL driver independent build passed (393.577s), isolated package hash ea1e3f3ac17e52e6 signed/verified. Actual Dev registry/factories loaded all six; new MySQL runner performed real bound Unicode/NUL/composite-PK grouped update, defaults/NULL insert/delete and full rollback through existing Mac wrappers. Installed Mac/Dev library unchanged.

• Final phone/iPad native forms inspected; both simulators now show the owned editable example table (querycraft_ios_edit.edit_rows) with safety lock on. Writable account has only SELECT/INSERT/UPDATE/DELETE on this owned DB, previous read-only fixture is intact. Mac integration writes were restricted to separate owned mac_* tables. All transient containers/volumes cleaned, only retained experience remains. Scope, unavailable editing surfaces, uncertain-commit behavior, engine/PK/binary/placeholder limitations and driver host dependency recorded. Final whitespace/shell checks passed. No commit/tag/release, no physical install or actual desktop input injection.

SQL query result editing — 2026-10-22

Implemented shared actual Mac origin resolver/query preparation and mobile MySQL single-relation update/delete using existing staging/native Forms/SQL preview/safetyLock/transaction service. Captured result/configuration, metadata cancellation/stale fences, query draft/history identity, original raw PK patches, duplicate projection overlays, committed-only export and navigation/connection/pending guards. Shared Mac query inline/inspector/delete path rejects expression-as-PK and conflicting/null physical keys; ordinary table behavior preserved. Complex queries, record/fullscreen editing and arbitrary write SQL remain explicitly unported.

Full phone/iPad each 239/0/0; focused final native/state checks recorded separately. Mac build and actual query TLS origins/delete/resident-patch/schema checks plus six actual registry factories passed. Signed ARM64 verified; physical phone disconnected/unavailable, not installed. Native query/form/pending screenshots inspected. First blocked-configuration recursion and unquoted Unicode-render limitations preserved as failure evidence; final guarded restoration/quoted-alias/native checks passed. Docs and owned live experience updated; transient containers removed, retained experience alive. No desktop mouse/keyboard injection, installed Mac/Dev/driver-library replacement, commit/tag/release.

Fullscreen and record editing — 2026-10-22

Complete: traced Mac coordinator/inspector/common requests, added one UI source adapter and immutable/generation-validated presentation capture reused by ordinary/fullscreen/record modes. Existing common staging/PK/type/preview/safety-lock/transaction semantics retained. Fullscreen tap edit/inspect, live pending colors, transient long-press/accessibility preview, actual controller/responder-aware keyboard protection, keyboard-stable viewport and restore. Record detail all-field edit/inspect/delete/restore/preview, large native sheet, retained independent record scroll and grid width/position/selection; stale source/raw replacement rejects edits and closes old detail.

Phone/iPad each full 248/0/0, focused native/state 16/0/0 each; phone large-layout follow-up 4/0/0, forced actual scroll-view recreation 1/0/0 each. Native table and real retained MySQL aliased-query full/edit/record/preview renders inspected on both. Earlier native guard/offset failures recorded; corrected actual keyboard viewport behavior and invalid test offset. ARM64 signing/strict verify passed; phone disconnected, not installed. Mobile-only scope and six-driver limits documented; no Mac/external-driver modification or rebuild claim. Transient MySQL containers/volumes cleaned; retained writable owned fixture and simulator experiences kept with safety lock on. No user desktop input, commit/tag/release.

2026-10-22 — Mobile connection/background reliability complete

Traced the actual Mac refresh/reconnect chain and retained mobile's fresh-per-request connections. Added explicit retry feedback in existing fixed status areas; background cancellation invalidates reads without clearing results/drafts, relocks and cancels writes, and uses bounded OS time for transaction cleanup. Rollback bypasses UI authorization; COMMIT invocation failures are uncertain and never replayed. Unknown-result discard remains readonly until successful verification, including failed refresh. Confirmed commits finishing in background also require an explicit table reload. Schema/export work pauses; completed Files input survives.

Final phone .build/reliability-phone-final.json: 262/0/0, 49.511 s; final iPad .build/reliability-ipad-final.json: 262/0/0, 49.744 s. Four new owned real-transaction faults, stale/background/read retry/draft/metadata/export state tests, native grid geometry/selection and existing regressions passed. Initial compile and standalone-host scene environment failures are recorded in findings/VALIDATION; not counted as acceptance. Final signed ARM64 .build/reliability-device-final.json: success, 22.346 s, strict deep codesign passed. Physical phone disconnected, not installed. Simulators restored to retained owned experience; transient fixtures removed. No desktop input, driver/Mac changes, commit/tag/release or persistent workspace expansion.

2026-10-22 — SQL selection/current statement complete

Shared the exact Mac execution target/resolver/preflight/plan sources and 24 original tests. Mobile captures actual UIKit selection/IME and source revisions; asynchronous validation rejects changed/cancelled/background contexts before connecting. Results, history and editing metadata use actual executed SQL, and public refresh repeats the result/failed attempt independently of later draft/caret changes. Editor text/selection stay in one workspace; selected batches are explicitly refused, never partially run.

Final phone 302/0/0 (44.046 s), iPad 302/0/0 (49.880 s); new native/real MySQL checks passed. Mac full build 56.001 s, actual compiled three-SQL-driver target/preflight and unchanged full-document batch rules passed; six isolated driver factories loaded. ARM64 17.587 s + strict deep signature passed, phone disconnected so no physical install. Both simulator three-statement experiences rendered and inspected. Initial compiler/parser-revision/test-fixture/resource-path failures are recorded honestly in findings/VALIDATION. No user input simulation, driver packaging, commit/tag/release or user database writes.

2026-10-23 — Six MySQL table operations complete

Implemented rename/drop/truncate/DELETE-clear/reset-auto/set-auto using the actual shared Mac statement, input/copy/capability models. Native catalog context menu and existing table/structure menus open compact confirmation with SQL preview, exact table name, same safety lock and actual current auto-increment metadata. Independent command execution preserves explicit DELETE/TRUNCATE semantics, checks target pending edits and captured context twice, updates affected catalog/navigation/query metadata, and fences completed/unknown writes across failed reads, dismissal and backgrounding.

Final full phone/iPad each 324/0/0 (46.238 / 57.002 s), including owned real MySQL six-operation semantics, permissions/FK/no-auto, lost-acknowledgement protection and actual App rename/drop route. Native forms/Return keyboard checks and shared six-action tests pass. Mac complete build 136.968 s, actual compiled six-driver capabilities/SQL and six original bundle activation/session factories passed. ARM64 build 36.617 s and strict deep signature passed; physical phone disconnected, not installed. Earlier compiler/identity/test-event failures are documented, not counted as passes.

Both simulator native previews inspected and restored to the retained owned writable action_demo experience with safety lock on. Transient fixtures cleaned; no user desktop input, arbitrary write-SQL expansion, external driver replacement, user database mutation, commit/tag/release. Mobile remains MySQL only; other five drivers evaluated and existing desktop behavior preserved.

Physical iPad installation follow-up (2026-10-23)

The paired iPad Pro (iPadOS 26.5) is connected/available after enabling Developer Mode. Initial install failed because the existing embedded provisioning profile did not include this new device. Rebuilt for the named physical iPad with automatic signing and device registration: .build/ipad-physical-signed-build.json succeeded in 14.887 s, no diagnostics; strict deep signature verification passed. .build/ipad-physical-install-final.json confirms successful installation. .build/ipad-physical-launch.json was denied by iPadOS Security, reporting signature/entitlement/profile trust; device-side Developer App trust is still required before launch. No successful physical launch or touch acceptance claimed. No proof environment or test DB credentials supplied to the physical launch.

Physical iPad follow-up: user confirmed Developer App trust. .build/ipad-physical-launch-trusted.json now reports SUCCEEDED, PID 661, no diagnostics. Installed current MySQL test app launched with no proof environment or injected connection. Physical touch/layout/database acceptance remains user-driven; no screenshot/UI verification claimed.

Physical iPhone update (2026-10-23): device connected on iOS 27.2; .build/iphone-physical-install-current.json and .build/iphone-physical-launch-current.json both report SUCCEEDED with no diagnostics. Same current signed build as the physical iPad, launched without proof environment/injected connections. This confirms installation/process launch only; no physical UI/database acceptance claimed.

Mobile display name (2026-10-23)

Changed the shared iPhone/iPad display name in XcodeGen from QueryCraft MySQL to QueryCraft; app bundle identity and data-container identity retained. Regenerated project, signed ARM64 build passed in 12.624 s with no diagnostics. Built Info.plist display name and bundle ID asserted, strict deep signature passed. Both physical device installs and launches succeeded (.build/mobile-display-name-{phone,ipad}-{install,launch}.json). Metadata-only change; no new behavior tests or driver changes, no claim of additional mobile driver support. No uninstall, user-data migration, commit or release.

Schema editing started — 2026-10-23

Authorized field/index editing plus table/list presentation. Reading actual Mac chain first; use existing worktree and preserve previous mobile and desktop changes.

Schema intermediate verification: full phone 343/0/0 (62.103 s); native 340-case predecessor also passed but wrapper failed due a mid-run script edit (now frozen and bash syntax valid). Mac complete build succeeded 385.076 s; signed third-party stripping warnings only. New MySQL driver is rebuilding separately. Final small presentation corrections preserve list row identities/selection over metadata refresh, make SQL preview rows immutable, and explicitly report/close confirmed success after reload. Final native checks and iPad full suite must include these corrections.

Recovery in continuation chat — 2026-10-23

Original worktree was accidentally archived during handoff. Recovered all tracked/untracked source through archive snapshot 3195231 in /Users/weilai/.codex/worktrees/ios-ipad-continue/QueryCraftPublic, attached to this chat. Original chat unarchived without restarting its stalled turn. Ignored .build outputs were removed by archival, so previous reports are historical evidence only; rebuild final acceptance here. Remaining user requests include iPad sidebar darkness and macOS Settings darkness. Physical iPad connected, iPhone currently disconnected. Main checkout's unrelated SSH changes untouched.

Continuation verification: full phone 343/0/0 (176.766 s); initial iPad 342/1/0 (215.650 s) found asynchronous SQL completion display invalidation, fixed by clearing text/caret-mismatched candidates immediately and preserving same-context candidates. Native iPad sidebar canvases and schema column/grid/list images inspected; Mac general/appearance/query/data light/dark windows captured independently via ScreenCaptureKit, no desktop input or activation. Mac full build 582.420 s, signed ARM64 build 149.519 s and strict deep signature passed. Final suites must include the completion fix and modern trait initializer.

Schema editing and appearance complete - 2026-10-23

Final normal-source suites: phone 343/0/0 (82.983 s), iPad 343/0/0 (133.231 s). New MySQL build 494.541 s, isolated package/signature and actual Dev API/registry TLS metadata/generated-plan smoke passed; original and new-MySQL mixed six-driver factories passed. Mac build 582.420 s and eight actual Settings light/dark renders passed. iPad/sidebar and both native table/list/column/index experiences inspected; DEBUG render ordering corrected before capturing true index sheets. Full tests preceded that fixture-only reorder; final simulator build 35.557 s and final signed ARM64 20.203 s passed afterwards.

Both final physical installs succeeded and strict deep signature passed. Final iPhone launch PID 4206, iPad PID 988; no injected environment/profile or uninstall. Physical touch/IME/rotation/database acceptance not claimed. Both simulators retain owned live action_demo schema experience in dark mode, default safety lock and no drafts; temporary test containers removed, only querycraft-ios-schema-experience remains. Detailed initial failures and wrapper-vs-test/render distinction recorded in VALIDATION. Source worktree attached here, original chat unarchived, main SSH work untouched. No feature commit/tag/release.

Started iPad completion visibility fix in the recovered worktree. Existing uncommitted schema/appearance changes retained; no branch switch or release action.

iPad SQL completion visibility completed - 2026-10-23

Replaced bottom iPad accessory with native caret-anchored table overlay; vertical scroll, two-line labels/type details, keyboard avoidance, native selection/touch/Return/Tab and lifecycle cleanup. Mac shared parser/catalog/ranking/source guards remain; phone accessory preserved. Actual Mac caret/window behavior traced before implementing iPad presentation.

Final full iPad 345/0/0 (225.858 s), phone 345/0/0 (136.464 s), no diagnostics. Native new tests cover reachable last row, acceptance dispatch, dismissal/focus and window/column/keyboard bounds under portrait/landscape/narrow/no/docked/floating layouts. Real owned MySQL SQL editor render inspected at ios-validation/ipad-completion-caret.jpg; candidates at caret. ARM64 33.787 s and deep strict signature passed. Device disconnected before installation; user explicitly selected simulator delivery, so no device success claimed. Simulator retained on owned live SQL completion scene. No desktop input or user database mutation.

User wants an on-screen keyboard and finds below-SQL candidates uncomfortable. Prior live render only exercised simulator hardware keyboard mode. Disabled Simulator hardware-keyboard connection via preferences and reopened Simulator without desktop input simulation, then captured actual software-keyboard scene before further layout decisions.

2026-10-23 SQL snippets complete for user-selected simulator delivery: shared Mac store/template/ranking/session plus native management and touch navigation. Corrected IME mutation replay using separately observed native text. Focused final 39/0/0; full phone/iPad each 380/0/0; Mac complete build and original snippet 56/0/0 passed. Final fixture-only snippet trigger rebuild/native presentation 2/0/0. Native phone/iPad light/dark navigation and manager/template renders inspected; actual live SQL scenes ready with sel candidates and software keyboards. Previous completion fix installed on phone but launch failed over connection; signed snippet build installed on iPad but launch denied while locked. User then selected simulator delivery, so no further physical actions or unlock waiting. No desktop input simulation, user DB changes or commit/tag/release.

2026-10-23 touch completion complete: adopted user-delegated keyboard-top shelf for docked iPad software keyboards, caret list for hardware/floating keyboards, and original phone accessory. Final .build/completion-touch-ipad-full.json 346/0/0 in 76.273 s; .build/completion-touch-phone-native.json focused editor suite 8/0/0 in 53.038 s; no diagnostics. Actual live keyboard/candidate render inspected at .build/ios-validation/ipad-completion-software-shelf.jpg. Simulator-only delivery per user request; latest shelf change is not installed on physical devices. No desktop input simulation, user database changes, commit/tag/release.

2026-10-23 iPad reconnect/install complete: paired iPad now connected/available. Latest keyboard-top touch shelf rebuilt/signed/installed/launched successfully in 23.069 s, PID 1146, no diagnostics; deep strict signature passed. Report .build/completion-touch-ipad-wireless-run.json. No proof launch environment, new source behavior, physical input simulation or iPhone update.

2026-10-23 completion acceptance/readability complete: native 13/0/0; final iPad full 351/0/0 (133.311 s) and phone focused 13/0/0 (92.582 s), no diagnostics. Actual opaque shelf software-keyboard renders inspected in light/dark. Signed ARM64 build/install/launch on physical iPad succeeded in 77.086 s (PID 1174), strict deep signature passed; report .build/completion-chain-ipad-device-run.json. User's standing preference is to deliver each verified mobile change to this iPad automatically; recorded in task plan. No physical input simulation, user DB changes or commit/tag/release.

SQL formatting complete — 2026-10-23

In the restored ios-ipad-review worktree on codex/ios-ipad-editor, shared the actual Mac formatter/worker/range/error/options sources and original formatter tests. Added existing-row mobile menu and Cmd-I, native grouped undo, composition guards, captured UTF-16 text/caret generations, cancellation on config/background/disappearance and teardown. Formatting changes SQL text only; standard uppercase/four-space defaults, no SQL execution expansion.

Initial resource-marker build failure corrected; focused 22/0/0, then final full owned-MySQL iPad 404/0/0 (237.007 s with rebuild), phone 404/0/0 (76.977 s). Eight native formatting checks and all existing regressions pass. Actual phone/iPad light/dark workspace attachments inspected; temporary DB fixtures cleaned. Retained iPad demo restored using the existing table-admin route after the old reader account was rejected. No user database/data/password changes, physical installation, desktop input simulation, external driver changes, commit or release. README/parity/validation reflect current snippet and formatting capabilities and remaining limitations.

SQL batch/results complete — 2026-10-23

Shared Mac target/preflight plus extracted SQLExecutionBatchRunner now drive bounded mobile current/selection/document execution. Same-session MySQL read-only batches prevalidate all SQL, retain independent typed results/editing metadata/drafts/viewports, stop on errors/cancellation/truncation/budget and mark untouched suffix. Native result menus preserve editor/grid geometry; public refresh retries only the chosen statement, actual successful completions write history, and all result drafts protect endpoint/table/schema operations.

Initial full iPad 419/1 exposed single-result row-loading versus editing-metadata lifecycle regression; corrected. Final iPad 420/0/0 in 97.012 s and phone 420/0/0 in 65.441 s, including owned real session/empty metadata/server failure/truncation/cancel checks. Actual phone/iPad light/dark renders inspected. Mac build 222.410 s and actual compiled common SQL runner regression passed; standalone SDK/C-module/resource harness setup failures documented. Desktop driver/API identities unchanged; no added other-five-driver live coverage. Latest simulator override remains; no physical installation, user database mutation, commit or release.

New MySQL table creation — 2026-10-23

Implemented Mac-aligned new-table creation in the isolated ios-ipad-review worktree. Extracted the unchanged WorkspaceNewTableDraft for shared compilation; existing native column/index editors now use one context for both existing and new tables. Actual server choices and all ten table options feed the same MySQL provider/builder. Catalog footer entry retains drafts; captured SQL preview and shared safety lock precede independent DDL, with final context/cancellation guards, successful navigation and refresh, and conservative unknown-ack fencing across close/reopen.

Original schema focused suite 11/0/0 and new state 10/0/0 passed. Owned real-MySQL full iPad 438/0/0 (116.901 s), phone 438/0/0 (74.242 s), no diagnostics. Five new real tests verify metadata/options/defaults/generated values/indexes, duplicate/denied failures, final authorization, actual committed CREATE with lost acknowledgement, and App catalog/search/navigation/empty-result behavior. Mac full build 63.159 s and actual compiled shared draft regression passed. Native phone/iPad footer, modal, definition, SQL lock and uncertain-state renders inspected. Extra light/advanced-scroll native verification final iPad 3/0/0 (39.230 s), phone 3/0/0 (29.307 s). Initial extra native 2/1 was a wrapper-vs-UITextField accessibility selector; corrected to native placeholder, failure retained/documented. Settled segment state and actual light/dark/advanced-scroll renders inspected; no application changes after full passes. Both simulators restored to the retained owned MySQL demo with new-table footer entry.

Simulator-only preference remains; no external driver/API changes, physical device action, user database mutations, commit/tag/release. Existing uncommitted SQL formatting/batch changes and unrelated main checkout preserved.

Existing MySQL table options — 2026-10-23

Implemented the Mac options chain on mobile: shared MySQL table-information query/conversion, existing schema state/options delta/provider/ALTER builder, one native new/existing options form and a menu entry in the existing structure footer. Options stage into the field/index draft; captured immutable preview and authorization fence stale plans. Explicit-field metadata comparison avoids replaying unchanged dynamic auto counters. Actual comment metadata propagates to sidebar, current object/data identity and route after successful reads. Loading and acknowledged/unknown DDL behavior use the existing chain.

Initial original schema 11/0/0; new options state 8/0/0 after fixing a Swift Testing non-Sendable key-path parameter. First full iPad 454/0/0 (246.013 s). Added stable one-time source/preview State and native parent reconstruction coverage; final iPad 454/0/0 (143.236 s), no diagnostics. Five new owned real-MySQL tests exercise engine/hints/default charset/counters/comments/data preservation/permissions/stale options/lost acknowledgement/navigation. Native final iPad light/dark/advanced/expired/SQL renders inspected; live retained owned schema demo restored.

Mac main build 74.154 s, independent MySQL build 280.443 s, local x86_64 package and strict signature passed. Actual Dev installer/registry loaded the new isolated MySQL package over TLS, read real options and generated a non-replayed-auto comment plan; all six real plugin activation/session factories passed. Initial harness Redis TLS mismatch corrected; user Dev installation/driver store unchanged. Final phone full 454/0/0 (91.979 s), no diagnostics; actual native light/dark/advanced/expired/SQL renders inspected and retained owned schema demo restored. README/parity/validation and task plan complete. No commit/tag/release, physical device install or user database writes; only disposable fixtures were mutated.

Start requested items 2 and 3 — 2026-10-23

User requests large-query streaming/disk caching plus arbitrary write SQL/transactions together. Working in the existing isolated iOS worktree, preserving all completed features. Applied planning-with-files-zh alongside existing SwiftUI/concurrency/testing workflow. Beginning shared store/transport/execution policy discovery; no feature implementation yet.

SQL cache/session implementation in progress

• Shared result store now has an additive resident-row threshold; mobile uses zero and compiles the same GRDB store and export row source. Mac default unchanged.
• Added typed transport query stream with 500-row/1 MiB batch backpressure, metadata/origins, binary preview, rowless affected-row acknowledgements and server transaction state. Mac MySQL query path adopts stream.
• Added mobile persistent SQL session, immutable write confirmation, safety-lock pre-send validation, transaction controls, cache pages and entire-cache export. First compile running; integration and cancellation lifecycle verification remain pending.
• First mobile builds: missing conditional Core import in shared validator fixed; regenerated project after adding confirmation source and made shared transaction command visible across MobileCore boundary. Third app build passed.
• First full test compile exposed Swift Testing macro accessing actor-isolated store let across test module; explicitly marked Sendable immutable store nonisolated. MySQL Mac driver V4 independent build passed (94 s).
• Found and fixed transaction command reusing the selected read-result identity, and parse failures falsely resetting active server transaction state. Commands now get their own correctly classified results; failure state is read from the actual session actor.
• Full iPad simulator run passed 462 tests (0 failed/skipped, 200 s), including new 100,000-row cache/export, immutable SQL confirmation, permissions, cross-execution session/savepoints/commit/rollback, background rollback, server implicit DDL commit, cancellation and typed binary/empty metadata checks.
• Added further tests for byte-bounded long-text pages/export, shared cache budgets, actual SQLite/WAL/SHM cleanup, interrupted-write verification fence and native iPhone/iPad cache/confirmation renders. Final full run now includes these.
• MySQL and Doris V4 builds passed. Packaging and isolated Dev registry load validation remain pending. Query result cache can also be explicitly cleared from existing SQL menu without an added toolbar row.
• Extended iPad run: 466 passed / 1 presentation test failed. All production cache/session integration, long-text bounds, cleanup and interrupted-write fences passed. Native test attempted execute immediately after changing SwiftUI binding, before UIKit received it; the actual editor correctly rejected stale captured text. Test now awaits native editor synchronization before execution; final rerun in progress.
• Isolated packaged-driver harness now preloads only old V3 framework before V4 (avoids unrealistically loading two copies of the Doris driver classes). MySQL typed 999-row query emitted multiple batches with NULL/binary semantics, and all six registry/session factories passed without runtime duplicate-class warnings.
• macOS test CLI rejected --arch (supported for macos build, not test); reran focused shared store/document/read-policy suites with native host architecture and supported arguments.
• Final iPad full run passed 467/0/0 (250 s); normal/dark cache-page and native write-sheet attachments exported. The sheet snapshot captured during transition, so strengthened render validation to wait for settled sheet and capture the entire visible window as well.
• Mac shared store/document/read-only validator suites passed 41/0/0 (312 s), including unchanged default resident/spill behavior. Package harness verified loaded V3 from installed old framework and V4 from the newly packaged MySQL bundle path, not an accidental derived-data transport override.
• Final audit found nil read retry could fall back to editor execution when an earlier write failed before replacing the prior visible query. Refresh now always uses the captured-read-only parsing path, including an initial empty result; added owned MySQL regression that creates the formerly missing table before refresh and proves no write is replayed.
• Final phone full regression passed 468/0/0 (124.070 s). Final iPad SQL/cache/native supplement passed 14/0/0 (47.599 s), including the new refresh no-replay regression.
• Settled iPad visible-window capture confirms complete SQL text, target database, caution and action inside the actual native sheet; no production rendering defect. Real phone app relaunch successfully removed the owned previous-run cache marker. Existing demo restored with safety lock enabled.
• Final metadata audit found a read begun before SQL write could publish after write invalidation; sqlDidWrite now cancels its existing read generation. Added a gated-state regression proving late reply cannot clear requiresReload, reusing existing SchemaStateService.
• Final phone focused suites 24/0/0 (44.315 s), final iPad schema state 10/0/0 (21.250 s). Latest cancellation/invalidation regression passed on both device simulator targets; both use x86_64 runtime on this host.
• Actual visible native confirmation windows checked on both devices after settling: full SQL, database, caution and explicit action visible; grid/cache pagination frames stable. Final app launch cleanup markers removed on both simulators, both restored to retained owned TLS schema fixture with safety lock on. No user desktop input, physical install, commit/tag/release, in-use driver replacement or main checkout changes. All requested items 2 and 3 complete; remaining limitations documented.

User-requested branch commit — 2026-10-23

Re-read release process and root instructions before staging. Worktree was detached at the exact existing codex/ios-ipad-editor tip, so switched back without modifying files or main checkout. Checked app version, Driver API 5 constants/manifest templates and v0.1.12 driver diff; MySQL/Doris require future paired V4 releases, other four driver releases are not warranted. Added bilingual pending release note with validations and explicit unfinished release-only gates. Existing full/focused structured reports remain successful; this commit changes no tested runtime logic.

2026-10-23 Redis full mobile acceptance (active)

User asked to finish Redis iPad and iOS before final delivery. Inventory underway; member search implementation is in workspace, not yet final. Sessions running: mobile 7.4 iPad search tests 17037; Mac Redis tests 5950; arm64 build 3939. Earlier baseline search build passed 33 existing tests, before adding 4 search tests and 1 UI search test. New expected scoped count 38. Continue checking actual results, do not count prior runs as final coverage.

Console source closure compiled into mobile project and model/sheet entry wired; simulator build running session 81022. Search old first 33 regression build passed; new 38 run had 1 score-text expectation failure, corrected. Mac Redis 11 tests passed and arm64 search build finished, but they predate console. Need rerun after console/whole Redis changes. No final delivery until active whole-Redis checklist is complete. Console first compile caught @Sendable executor accessing @MainActor controller state; moved execution handling into an isolated async controller method and made callback hop via await. No unchecked Sendable workaround. Console next compile hit a Swift compiler crash (SwiftCompile, no normal diagnostic; log build_sim_2026-10-23T14-18-13-653Z_pid18039_0796b8f8.log). Need inspect crash request and simplify closure initialization / macro isolation before rerun; do not repeat identical build without changing cause. Compiler crash is IRGen in MobileRedisConsoleView.swift, generated String @MainActor thunk. Replaced method-reference Binding setter document.updateSource with explicit closure to avoid Swift 6.2.4 IRGen bug; previous callback isolation fix is valid, crash was in view not controller. Current build session 65399, output .build/ios-validation/redis-console-build.json. Earlier sessions 81022 / 32509 / 47400 done or likely done; read result before new builds. Need inspect @MainActor Binding fix result. Last validated before console: 37/38 search test passed, exact score expectation corrected; no full Redis acceptance yet. Latest running iPad scoped test session 93778 now includes shared console/Stream/String plus new console/stream tests and native large String/console test (expected 43). This run predates last optional resource metadata + expired-type guard patch. Continue, do not finalize. Console background-held QUEUED test added: waits for actual callback then confirms closing discards unEXECed transaction. JSON toggling/lexical numbers, binary output bound, SELECT queued inside MULTI tracking tested. New typed console protocol allows deterministic lost-ack injection without executing user DB. Latest console test build (93778) failed on Swift Testing throwing-expression macro inference in a closure plus unused map argument. Replaced first fixture's throw-in-#expect expressions with explicit await result locals and supplied _ in; avoids inference issue and records actual values. This was compilation failure, not passed test evidence. Latest 65388 build failed Swift expression type-check timeout in resource details initializer after metadata additions. Broke columns/attributes/status into explicitly typed locals and removed redundant optional flatten. No runtime test result yet for the full-console set. Tests count expected 43. Runtime issue: native large String presentation test stalled for >2min at testResourceConsoleEntryAndLargeStringNativeEditor with 1.2MiB (100k lines). Do not ship giant UITextView full-value layout. Stopped own test/build CLI process trees (28733/29707) and own simulator app28964; two test runs accidentally overlapped same simulator/report, neither counts as valid. Need make large String editor bounded segmented UI while retaining full raw draft/CAS, then rerun sequentially. Old test session77723 / new66344 terminated; scripts should run finally and remove own Redis fixtures. Must inspect cleanup. Native large String UI corrected to bounded 32KiB UTF8-boundary-aware segments above64KiB. Full string draft retained; only selected byte-aligned segment replaced, full-value planner/CAS unchanged. Added explicit page controls and explanatory text; adjusted native test to assert bounded visible text plus complete saved length. Need add multi-byte boundary segment round-trip/conflict tests and rerun. No further same-simulator overlap: wait for running session completion before any new simulator build/test. Active work not accepted yet: selection/export and metadata actions were added after running session73627 started. Wait/check it, regenerate project, run new whole Redis scoped tests (~45 methods), then full suite. Need preserve sequential simulator execution; do not launch another simulator run while current one active. Extra tests for controller gates and UTF8 String segments now in MobileRedisConsoleTests.swift. Add latest test/result and failure repairs to this log. 73627 completed successfully: 44 passed/0 failed/0 skipped (64.400s), including native segmented >1MiB String, full saved length, console, shared search, UTF8 boundary/CAS-tail test. Predates new whole-key metadata actions and latest resource tree/export updates; do not call whole-Redis accepted yet. Next scoped run just started after xcodegen, expected45; Mac Redis tests and unsigned arm64 run independently (different derived data). 63325 succeeded 45/0/0 (151.524s); Mac driver37561 succeeded11/0/0 (141.864s); unsigned arm6454887 succeeded. These predate latest console memory retention/input segment utility changes; expected next scoped46 after retention blackbox test. Native1.2MiB segmented String no longer stalls; existing whole-value save length assertion passed. Whole-key metadata List/Stream/missing tests passed. Stage4 implemented but NOT accepted until final phone/tablet/full-suite/package/load and native previews. Current simulator scoped session18684 expected46 (includes retained-console budget test and native segmented input). Running independent Mac Feature pure regression selection (document/history/parsers/tree/plans/exports, excludes native event/window tests) in its own session, plus unsigned arm64. Do not launch a second simulator while18684 active; do not launch another Mac derived-data job while Feature regression active. Final full five-fixture mobile and packaged driver/Stream/CAS remains pending. 46/0/0 scoped iPad completed88.209s (18684); predated latest Hex codec/memory/worker fixes. HEX codec + binary command escaping now single buffer streaming rather than per-byte temporary String arrays (same exact formatting); avoids huge transient allocation for accepted100MiB values. Large String decode/format/validation is now worker-actor async with generation guards/cache; tests that change large drafts await prepare()?.value. Raw export and command export resolve complete current bytes asynchronously; JSON presentation resolves uncached edited large drafts, keeps raw bytes unchanged. Need rerun compile/tests for latest worker changes and inspect flags/cancellation. No simulator running now. Mac Feature19473 and arm6439548 may still running independently, inspect before reuse. Only one simulator run currently40254; it predates latest worker guard/cache/JSON, shared Hex/formatter optimizations and memory flags. Need await its result then regenerate/retest current. Mac Feature19473 likely finished; inspect result before next Mac job. Last accepted baseline46 succeeded; latest changes not yet validated. Active whole Redis goal remains pending final acceptance, not permission blocked. Whole Redis feature work now implemented; remaining is verification/fixes, no planned new feature scope. Touch header actions use 44pt icon targets for compact widths; Forms dismiss keyboard interactively. Last known baseline46+Mac28 passed, current source has streaming codecs/async cache/formatter fixes that need fresh builds. Need cover async HEX toggle and raw/command export of edited large string, unverified write/transaction cases already in ConsoleTests.

Latest shared formatter refactor caught two leftover integer index operations during compile; corrected UnicodeScalarView backward traversal. Fresh iPad Redis46/0/0 and Mac shared29/0/0 pass. Added deterministic idle-close/foreground transaction gate and native dark catalog/Stream TTL action tests; first full-run compile caught Swift Testing throwing macro as trailing closure expression, extracted actual reply before assertion. Rerunning full five-fixture suite. Current arm64 build succeeds55.704s, pending final test-source-only updates do not affect product.

Mac Redis latest11/0/0(187.209s), Mac shared29/0/0(86.245s), independent Redis package/signature completed. Final review found primary collection grid still used raw ZSCAN although editor used shared descending ZSet reader; replaced main List/Hash/Set/ZSet reads with same RedisCollectionReader, opaque type-bound continuation and lossless binary HEX display. Added real205-member grid order/next pages/List absolute index/binary Hash test. Full iPad run8827 may predate this last grid change; supplement afterward and phone full latest.

Full iPad owned five-fixture regression succeeded567/0/0(234.443s). Saved redis-complete-ipad-full.json; SQL rollback fixture emitted existing “there is no transaction in progress” warning, no errors/failures. Predates final primary-grid parity/disabled empty selection footer changes and third new test; require final scoped iPad supplement. Phone full current-source running serially in same simulator derived data.

Final console audit: verification PING had accepted a server .error reply as a completed verification. Corrected to require real PONG and propagate ACL errors; added real restricted ACL user + committed/lost-receipt proof that denied verification retains fence and never replays. Phone full current run predates this final local service change; final scoped7.4 both devices plus arm64 cover it. No further planned production edits.

Final whole Redis acceptance completed: full iPad567/0/0(234.443s), phone568/0/0(179.128s), latest-code Redis7.4 iPad50/0/0(81.139s) and phone50/0/0(60.427s) including ACL/read verification fence and primary-grid parity. Latest arm6413.098s passes; Mac shared29 and driver11; package ad4e7535116eb1cd actual installed registry harness validates all6 and live new Stream+TTL/delete. Native phone/tablet light/dark renders inspected; no physical mouse/key input or user DB mutations. Current README/MAC_PARITY/VALIDATION updated with true remaining platform/protocol boundaries. No commit/tag/release.

Post-test cleanup verified: only preexisting querycraft-ios-schema-experience Docker container remains. Both latest simulator apps relaunched without fixture/test launch environment (PIDs70991/71674), preserving saved libraries; preliminary stop returned “nothing to terminate” because test runner already stopped the apps. No launch error. Final diff whitespace check clean.

User-requested local commit — 2026-10-24

Re-read root AGENTS.md and docs/RELEASE_PROCESS.md. Verified feature branch, preserved separate main-checkout changes, inspected current accumulated multi-driver scope and v0.1.12 baseline. Added bilingual pending mobile-driver notes with app0.1.12(22), API5 and per-driver future release decisions; no version bump/push/tag/publication. Rechecked stored final successful five-fixture and Redis50-per-device/Mac/arm64/package-load evidence. Only new commit-preparation changes are documentation and trailing whitespace. Stage all185 authorized source/test/doc/script files from this worktree; no build artifacts or credentials.

Elasticsearch full mobile delivery — 2026-10-24

Started user-authorized complete ES parity work. Read root instructions and planning/concurrency/build skills; inspected previous complete Redis acceptance. Initial scope includes all current Mac ES entry points. No modifications to main checkout, user DBs or desktop input.

ES shared compilation baseline passed (es-shared-build2.json; no errors/warnings). Initial failure was missing portable clipboard error/parser dependencies; sharing original parser behind platform-specific pasteboard guards resolved it. New native workspace/action implementation is in progress, not built/accepted yet.

First native build caught ambiguous MobileGridData overload on empty arrays and an API-key label mismatch; renamed typed grid initializer and used exact authentication case. Added default-preserving optional response budget to shared ES session (mobile16MiB; Mac default unchanged) and reject timed-out/failed-shard pages rather than treating partial empty hits as success. Native console, management, filter and main view now implemented; validation pending.

Mac desktop build is running separately. Shared driver tests were adapted to compile identical transport tests on iOS using the direct driver configuration initializer (original test assertions unchanged). First test compile found desktop config wrapper dependencies; removed only those helper wrappers. ES-only disposable fixture harness and real management/PIT/console/conflict tests added. Staged per-field edits reuse original Mac cell editor/model; bulk document deletion uses per-document version guards.

First live-harness test compilation caught @MainActor inference on pure test root-response helper and environment trait accessor. Made those helpers explicitly nonisolated (no actor state accessed). Excluded Mac-only WorkspaceModel page-refresh suite from mobile source list; mobile and shared driver paging tested via native model/real PIT instead. No runtime test result accepted yet.

First focused shared/mobile state run succeeded: 38 passed/0 failed/0 skipped (127.891s). It predates final native/live additions and later lifecycle/count guard updates, so requires latest full ES acceptance. Mac app build succeeded211.363s (existing Keychain/Sparkle warnings only); unsigned arm64 build succeeded103.957s. Real disposable ES run is now compiling/running all shared driver suites plus full new live/native/state cases.

Mac shared ES regression passed163/0/0 (133.455s), covering parser/DSL/management/document/cell/paste/template models. Current simulator live run and separate driver tests in progress. Last UI audit added selection application guard to native REST editor and automatic confirmation sheet for keyboard execution; source-generation guards remain. Pending field display is separated from committed-data export, so exports never include unsent drafts.

Real ES acceptance exposed a genuine lifecycle crash: desktop refresh replaces sessions, while mobile had called close then connect on the same URLSession-backed driver. Fixed by sharing a read-context reset method (PIT/mapping/scores only); actual cancellation/background close now drops the actor and uses a new factory-created session after awaiting shutdown. Desktop behavior unchanged. Native fixture also corrected integer mapping to long for the intentional9007199254740993 sample. Failed/crashed run is NOT validation evidence.

ES iPad first complete scoped acceptance passed74/0/1 (80.105s), including actual native management screens and deep PIT/conflict/console/admin/state tests. One original driver live test requires its dedicated fixed-port environment and is skipped; equivalent mobile-owned real mapping cases passed. Subsequent final-source additions are complete-match export via independent shared paged source, read-drain before shutdown, and shared response diagnostics/error navigation/task followup. Five-fixture full iPad now running; final source must be regenerated and retested afterward if this run predates new files. Driver independent build passed25.244s and signed local package created.

Final arm64 compile caught SwiftUI type inference ambiguity in the document view after adding optional Double score display; replaced overloaded String.init method reference with explicit typed conversion closure. Re-run required. Standalone package-load harness initially lacked transitive Clang module paths; switched to the same modulemap/include set as repository native SQL harness. Package itself built and signed successfully; loader has not passed yet.

Packaged ES actual installer/registry load passed all six driver factories; newly packaged e5a2f9c9fe3030a3 ES bundle verified typed PIT columns, full doc/routing/CAS and Mapping updates against disposable real8.18 server. Final arm64 build passed34.086s before final UI label/link polishing. Native iPad dark workspace, light confirmation/REST and dark Alias screenshots inspected. Final full iPad is in runtime; no test failure identified so far.

Final audit added explicit acknowledged receipts for partial document batches (no pretending rollback/retry), retained unknown-outcome fences, full-source-vs-preview share labels and human-readable resource labels. Native form fields now keep visible labels. Added real Data Stream document/backing-index test, diagnostic source-generation test and partial-batch refusal/replay test; final iPhone full run and latest iPad ES supplement must include these. No further feature expansion planned.

Full iPad run exposed an unowned-owner crash after background guards completed: a completion Task scheduled after workspace lifetime ended could read owner before cancellation checking. Fixed every action/console task to capture the workspace for its bounded task lifetime (released on completion), guard cancellation/background before completion work, and capture owner in keyboard/button async bridges. No permanent owner back-reference introduced; full run must be repeated and is not accepted even if remaining tests pass after runner restart.

Lifecycle audit also stores/cancels the simulation read task and drains completion work explicitly; action/console tasks keep owners only through execution. Phone workspace now restores selected detail/console on reopen, like existing Resource workspace. Product scope now frozen; only test-found fixes may follow. Latest ES scoped compiling; next required runs are fresh full five-fixture iPad + iPhone and final arm64.

Latest lifecycle-focused scoped run passed77/0/1 (93.430s). Final review identified interrupted staged-field validation could retain an old closed session; preview now rebuilds the same shared cell edits against current session while preserving original snapshot/CAS version, using a default-false forceValidation parameter (Mac default unchanged). Added real background+staged-resume regression. Full iPhone fixture startup active; final full tablet still required. Product implementation now frozen.

Final Mac console/response diagnostics supplement passed26/0/0 (296.832s; existing dependency/Keychain warnings). Latest host cell-change default-path regression and final arm64 rebuild running separately; phone full five-fixture runtime still active. No production edits after target-connection labels; only validation/documentation ongoing.

Full five-fixture iPhone run succeeded647/0/1 (329.766s), including latest staged-resume test. No fatal lifetime error. Final ES-only UI polish after that source: target connection labels, actual protocol top-level field menu (no dotted-name hierarchy guessing), and fixed-height response header with multi-result menu instead of an inserted toolbar row. Native batch result capture added. Fresh full iPad and final scoped iPhone must verify these; no changes to other drivers.

Final surface review found fullscreen ES field staging lacked the existing grid long-press pending-changes entry. Connected the same optional pending count/action to the shared fullscreen grid, with nil defaults preserving SQL/Redis behavior; fullscreen sheet now pauses viewport tracking while editing. Native ES test mounts fullscreen and opens the shared pending plan on both device targets. Final scoped two-device acceptance is required after full iPad.

Final template UI inventory found Mac searches both names and index patterns. Extracted its exact predicate onto the shared snapshot and used it from both Mac editor and mobile template list; added native search input and pure mobile matching regression. No modified matching/ranking rule. Final scoped suites and targeted Mac template tests must include this last parity correction.

Final five-fixture iPad succeeded647/0/1(295.434s), matching phone647/0/1. Latest-source phone ES acceptance79/0/1(95.607s) and iPad ES plus original fullscreen83/0/1(68.644s) passed. Native final phone/tablet template search, batch console and fullscreen captures inspected. Latest arm64 succeeded32.953s without diagnostics; Mac template12/0/0 and actual packaged six-driver/live new ES loading passed. Only final phone original-grid supplement, normal simulator relaunch and documentation/cleanup remain; no product edits planned. New harness removed nonexistent Mac-only suite and unused all-mobile option (five-fixture full regression uses its original harness); executable Python syntax checked.

Final phone original fullscreen supplement4/0/0(16.376s) passed with no diagnostics. Updated acceptance/validation/release-scope docs and marked ES task complete. All owned temporary containers removed; only preexisting querycraft-ios-schema-experience remains. No application source changes after final successful two-device ES/arm64 validation. iPad prelaunch stop reported nothing to terminate (test runner already stopped it), not a product failure; normal no-fixture-env relaunches requested for both simulators. Main checkout retains exactly its preexisting AGENTS/planning changes. No commit/push/tag/release or physical install.

Normal launches succeeded without fixture environment: iPad PID22181 and iPhone PID22180, no launch diagnostics. Saved libraries preserved. Final diff whitespace check clean; worktree contains only intended81 source/test/script/document paths.

PostgreSQL complete mobile delivery — 2026-10-24

Started user-authorized PG work in existing ios-drivers worktree, branch codex/ios-ipad-editor. Read root instructions, current planning and SwiftUI/planning skills. Existing ES source/tests/docs remain. No agents or desktop input; inspect actual Mac entry points before implementing.

PG implementation now shares actual schema provider/descriptor/executor and extracted exact index metadata/DDL conversion with Mac driver. Query-result target lookup uses protocol Schema/table independently and existing shared editor accepts canonical quoted identities. Shared object/new-table identity helper preserves old ordinary names and safely decodes quoted dot/quote names; existing driver fallback now uses same helper. Native forms use provider-specific choices and transaction semantics. View-edit helper briefly stopped because its first visibility replacement had matched both controls; corrected index-only descriptor flag and continued, not a product execution failure.

Initial builds caught a protocol method declaration accidentally inserted in the concrete MySQL service and an over-complex optional view-definition expression; corrected both. Sharing import sources next exposed AppCopy's mobile module boundary; added the same conditional MobileCore import used elsewhere. File-import transactional engine is now shared through a connection adapter; Mac target wrapper retains original strict-MySQL verification and execution order. Mobile PG adapter additionally locks/revalidates physical table inside its import transaction. Native import picker/options/mapping/preview/execute/unknown-outcome UI implemented and workspace write/lifecycle gates connected. COPY is already compiled LibPQ transport; Mac table file-import uses INSERT batches, while native COPY belongs to generic cross-connection transfer workbench, a separate workspace capability.

Second import build crashed Swift IR generation in an actor-isolated format setter method-reference conversion (WorkspaceDataImportFormat -> isolated Binding setter), with no source diagnostics. Replaced the method reference by an explicit MainActor closure; rebuild next, not counted as acceptance.

PG mobile build passed50.759s without diagnostics. Initial focused test compile needs original WorkspaceSQLPreviewStatement used by original provider tests; share it and its PG preview builder rather than weakening assertions. New PG fixture SQL shared verbatim by full/focused harnesses. Mac shared regression running separately. Documentation helper stdin encoding failed before writing; created parity matrix through apply_patch.

Initial native PG test assertions explicitly expected schema/result editing to be unavailable. Updated only these capability expectations to the new intended true state; all original behavioral assertions retained. Import preview now tracks the exact loaded options and invalidates acceptance on reload/failure, retaining old content with an explicit stale-preview label.

Second focused compile reached original XLSX test but Bundle.module is unavailable in native XCTest. Reused already-included original Fixtures folder with a test-bundle marker on iOS; parser assertions unchanged. Added native PG structure/field/index/new-table/import/query-pending/table-action renders and import state/identity/transaction rollback tests.

Mac shared regression55/3/0 caught real adapter-order drift: factory creation preceded validation/authorization, unlike original behavior. Restored preflight before session creation; generic engine still rechecks before connecting. Also narrowed explicit-name quoting to schemas that actually need escaping; ordinary schema + dotted table retains original audit.new.name presentation/selection while physical components still round-trip. All three failures require a fresh shared Mac run; not accepted.

Native test compile required explicit MainActor isolation on the async schema-body callback and its nested generic UI renderer; added annotations rather than moving UIKit work off actor. No runtime acceptance yet; product mobile build already succeeds.

Mac preflight wrapper requires await for its MainActor authorization closure; corrected compile error before rerun. Shared PG SQL preview now derives physical parts from the same computed object identity, keeping special-name preview aligned with actual execution.

First expanded phone compile caught an async test actor getter inside the intentionally synchronous MainActor authorization closure; made the test gate MainActor-bound. PG package scheme is build-only, so its test action did not run; resolve the actual driver test scheme before retry. Arm64 passed68.848s before the final large-cache editing supplement. Large PG cache pages now reuse shared result-store apply/remove; disk deletion compacts one payload at a time instead of fetching the entire cache into RAM. Need dedicated cache edit/delete/export and Mac-store regression before acceptance.

Scope audit found PG cached results over500 rows still disabled editing. Extended current completed auto-commit PG cache pages through the same editing model, updating/deleting raw rows in the same store and retaining whole-cache export. Added1501-row real edit/delete/page/export regression; asynchronous patch has explicit write/context gates and completion waiting. Store deletion now streams compaction in SQLite; Mac original resident/disk tests required. Original SQL transaction and incomplete-result restrictions remain explicit. No other driver gets newly enabled cache editing.

Cache metadata audit corrected cross-result invalidation to inspect preview.workspaceColumns (which retains PG Schema), not C result origins that omit Schema. Added real generated-column/index/comment new-table coverage and native scrolled import/new-table confirmation captures. IDENTITY label now stays generic for existing ALWAYS/BY DEFAULT modes. New-table final authorization also rechecks captured Schema. Product scope frozen again; only validation findings may follow.

Importer now captures the immutable workspace connection configuration rather than reading a mutable SQL editor password. PG-only harness includes the separate shared import-engine fault suite as well as native/state/real tests. No new feature work remains.

Final confirmation audit added physical Schema to PG table actions, schema pending target and new-table confirmation. PG sequence UI no longer labels an unavailable multi-sequence value as a single next number; it states owned sequences are resolved at execution. Existing MySQL rows remain conditional defaults. Final full iPad fixture startup is active; latest native two-device supplement must include these target labels.

Final cache lifecycle review found page changes reused resultID, so a retained editor from a prior cache page could remain authorized. Each page now gets a fresh editing identity, and mutation context rejects the optional metadata/page-loading interval. Added an old-page-sheet rejection check to the1501-row regression. Full iPad may predate this last guard; final PG iPad supplement and full phone must include it.

Final provider comparison found Mac explicitly supports zero-column PG tables while shared mobile MySQL validity required a field. Kept that MySQL requirement but allow PG zero-column plans, verify relation existence/kind when columns are empty, and add real create/load/add-first-field acceptance. This is the last baseline-specific correction; final PG supplements must include it.

Full five-fixture iPad succeeded694/0/1(397.236s), no errors/failures, existing SQL fixture warning only. It predates final cache-page identity/zero-column guards; full phone and final focused tablet supplement will cover latest source. Actual initial PG package-load failed because PackageFrameworks/QueryCraftPostgreSQLDriver was still the old Oct23 dynamic product: driver tests rebuild their test module, not the distributable dynamic framework. Independent product build is now running; must await it, repackage and rerun actual load. No failed package counted as acceptance and installed user Dev untouched.

Native import rows now have stable file-snapshot/row identities and source column names are stable ForEach IDs; old preview identity remains stable during refresh/failure. No business behavior change. Final source changes now limited to validation-found fixes and this native identity correction; do not add further capability scope.

The shared PG details already generated the same metadata-derived structure SQL as Mac, but native structure menu had no inspection entry. Added read-only captured-snapshot view/share in existing PG menu, labelled with its real field-definition scope; no duplicate refresh or other-driver entry. Final native two-device supplement will mount it. All source additions now accounted for in the parity matrix.

Fresh dynamic PG build198.502s succeeded; repackage/signature850f7b8725f6b4f7 and actual installer/registry six factories + new quoted Schema/table provider/index metadata/large integer live test passed (pg-packaged-load2.log). Latest arm64 pg-delivery-arm64.json succeeded32.122s without diagnostics. Full phone runtime active in pg-final-full-iphone.log / test_sim_2026-10-24T04-40-12-149Z_pid89495_0ae51711.log. Required remaining: save full phone report, final current-source PG scoped tablet then phone serially (includes old-page guard, zero-column, stable import rows, DDL/native target labels), inspect final images; final docs/clean normal launches. No further feature expansion.

Full phone695/0/1(299.480s) succeeded with existing warnings only. Final PG current-source tablet supplement running in pg-delivery-ipad.log (session34093), then phone supplement required. Fresh PG package850f7b8725f6b4f7 actually loaded all six factories and live special-name/schema metadata/large integer checks passed. Current-source arm6432.122s passes. README COPY wording narrowed to the actual COPY FROM real-test evidence; FROM/TO code sharing is not two live protocol tests.

Final tablet PG80/0/0(83.424s) passed with existing test/SQL warnings only. Latest tablet new-table/table-action/structure-SQL target captures inspected. Final phone PG supplement running in pg-delivery-iphone.log (session75072). Read-only final report check initially used main checkout and found no local report; corrected path, no writes there. Main status remains only preexisting AGENTS/planning changes. New-file source/test/script/SQL-fixture/doc inventory and whitespace checks clean.

Final phone PG80/0/0(41.799s) passed. Native latest phone import/new-table confirmation and query pending screenshots inspected; all actual target names and requests readable. Updated matrix/validation/README/release scope and task completion. Docker only preexisting querycraft-ios-schema-experience remains. Final current-source normal simulator launches pending tool-result confirmation; no more product source changes or tests required.

Final normal simulator launches succeeded without diagnostics: iPad PID97821 and iPhone PID97822, current-source MobileProof app. Launch reports saved as pg-delivery-launch-ipad.json / pg-delivery-launch-iphone.json. Delivery verification complete; no further product edits.

2026-10-24 user requested local commit and stop. Reread AGENTS.md/RELEASE_PROCESS.md; checked local v0.1.12 baseline, app0.1.12(22), API5 and all six cumulative driver decisions. Preserve existing verified test/package evidence; no source edits or additional feature work. Commit accumulated ES/PG sources/tests/docs only; no push/tag/release or installed-Dev changes.

2026-10-25 resumed user-authorized four remaining areas. Read current AGENTS and planning/SwiftUI/concurrency/XcodeBuildMCP skills; verified clean existing branch. Discovery ongoing before source edits.

Doris implementation: original Mac index conversion moved to same shared metadata enum, native structure adds actual SHOW CREATE/SHOW INDEX with explicit unsupported state and unified refresh. Mobile SQL joins same streamed disk cache but validates entire read-only batch before connect and sends no MySQL BEGIN/ROLLBACK/SET; transaction UI stays disabled. Mac Doris buffered query replaced by same backpressured transport with retained row-limit semantics. Validation pending.

First Doris compile succeeded, but test invocation used wrong target QueryCraftMobileTests; actual target is QueryCraftMobileProofTests. Report had0 executed and is NOT acceptance. Correct target for all follow-up test runs. Current CLI exposes no session-default tool; inspected workspace daemon state and will pass explicit project/scheme/device/derived defaults.

Initial native workspace regression passed23/0/3 (environment live fixtures absent) in59.622s. Shared document/archive UI and settings expanded after that baseline; final current-source tests still required. Transfer implementation now compiles original Mac planner/catalog/runner/manager and exact SQL driver sessions through conditional imports; no duplicate planner, SQL generation, COPY/LOCAL fallback, or transaction engine.

First transfer/shared compilation found missing iOS imports for SQL overview and absent original DatabaseConnectionConfiguration/DatabaseProduct sources. Added exact shared source files to Core and platform import only; removed accidental desktop task-tree UI source (native adapter uses original library backend). No business fallback or reduced transfer engine.

Shared transfer SQL/runner/library/bulk and mobile document regression46/0/0 passed96.068s. This predates final browser-state recovery and write-ahead archive fence; these require new coverage. Native settings use exact theme archive/roles, shared formatting and indentation/completion rules, visible-line gutter/background/selection decoration; UIKit keeps marked input/undo and native find (find colors explicitly system-rendered).

Owned workspace fixture first attempt failed during MySQL fixture creation because EMPTY is reserved and the synthetic field was unquoted. Fixed fixture quoting (retains physical field name), all owned containers removed by finally. No tests counted. Latest shared source adds task-tree/name-matching extraction for Mac/native reuse and native variable-width font/IME checks.

Owned workspace second attempt seeded MySQL/PG/Doris successfully; current iPad acceptance active (session97422, workspace-owned-ipad.log). Parallel Mac shared/core regression active in separate derived data (session25924, remaining-mac-shared.json), no Mac UI-input simulation. Final native test source includes font/theme/tab/IME, both device layouts and full migration confirmation/screenshots. Extra scope audit: verify Doris cached SQL comments follow actual protocol origin, alongside table structure, to avoid silently losing metadata on streamed results.

Owned workspace fixture reached all three ready servers, then compilation failed only on MainActor static test argument enumeration. Marked immutable Sendable pairs nonisolated; no runtime tests counted for that attempt. Extra Doris audit found SQL comments previously depended on writable single-table editing. Extracted exact Mac physical-origin mapping to shared helper used by Mac grid and mobile streamed MySQL/Doris metadata loading; aliases/joins/empty/expressions remain origin-based. Final tests must include this shared change and Mac inspector regressions.

First Mac shared run ended with0 tests: new shared metadata source was created after Xcode started enumerating package sources, so that in-progress build could not see the new helper. Re-enumerate source tree before current-source rerun and freeze file additions during validation. First owned iPad runtime57/0/0 passed before metadata/recovery/theme initial-appearance validation fixes. No failed/zero-test run counted. Remaining product implementation is now frozen except concrete regression fixes; finish validation/docs.

Final hardening before validation freeze: corrupt workspace can now explicitly back up the original then restore saving, with backup export reference retained across restart; saved-query deletion and no artificial32-document cap. Shared desktop/mobile color undo coalesces continuous picker updates identically. Physical-origin metadata shared across all3 SQL drivers; full type from real column metadata preferred. Write marker persists only when its acknowledged state changes, avoids repeated16MiB archive writes during a multi-statement preflight and restores prior marker after disk failure. No new capability scope remains. Final owned tablet (session94733) and Mac rerun(session57942) active.

2026-10-25 final workspace delivery: Doris, persistent independent SQL documents, original generic transfer and editor settings complete. Both full752/0/9; current-source80/0/0 each; Mac65/0/0 plus Doris5/0/0; three independently built/signed SQL packages and actual six-driver/liveDoris loading; latest unsignedarm64 and normal simulator launches passed. Corrected long-text scrolling, history context, namespace invalidation, durable/cross-document unknown-write fences and fixed native footer/confirmation layouts. All evidence and exact scope in QueryCraftMobile/VALIDATION.md and WORKSPACE_PARITY.md. Main checkout/user Dev/databases untouched; no commit/push/release. Physical/network/TLS and missing Simulator Emoji font remain explicit verification limits.