test(regression): suite for session bugs (MariaDB type names, RowDisplayCache, PluginCellValue.asText)

[datlechin]

Summary

Regression tests covering the testable units fixed during the recent debugging session. Every test corresponds to a bug surfaced and fixed in a prior PR; the goal is to catch the same regression if anyone refactors these paths again.

Coverage map

Bug / PR	Test suite	What it pins
#1209 numeric rendered as hex (0x31 for 1)	MariaDBFieldClassifierTests (existing) + new MariaDBTypeNameTests.numericTypes / decimalAndFloat / temporalTypes / miscTypes	type code → name mapping for INT family, DECIMAL, FLOAT, DOUBLE, all temporals
#1209 numeric routing — field.charsetnr == 63 flagging non-binary types	MariaDBFieldClassifierTests (existing)	classifier returns false for numerics/temporals even with binary charset
#1217 BINARY/VARBINARY classified as text → data wipe on click	MariaDBTypeNameTests.binaryWith* + varbinaryWith* + charWithout*	case 253 → VARBINARY only when binary-flag set and charset is 63; case 254 → BINARY likewise; non-binary stays VARCHAR/CHAR
#1216 mariadb_field_attr JSON over-read	(covered by MariaDBTypeNameTests.miscTypes checking type 245 → JSON; the over-read fix is in code paths that consult the function from mysqlTypeToString, which is integration-level)	type 245 returns JSON (the JSON detection from extended metadata requires real MariaDB connector and is integration-tested only)
#1212 P0 #2 — drop NSCache + RowIDKey for Dictionary	RowDisplayCacheTests (new)	basic set/get, FIFO count eviction, cost-cap eviction, replace-same-key doesn't double-consume queue slot, removeAll resets state, mixed RowID.existing / RowID.inserted round-trip
#1217 .bytes.asText returning empty would have caused click-to-wipe	PluginCellValueSortKeyTests.asText / bytesAsTextIsNil / nullAsText / asBytes (extended)	enum-case → text/bytes mapping is pinned so the editor can rely on it

Refactor for testability

mysqlTypeToString(_:) previously took UnsafePointer<MYSQL_FIELD>, which can't be constructed without a live MariaDB result handle. Extracted the pure switch into a new mariaDBTypeName(typeRaw:flags:charsetnr:length:) -> String (internal). The MYSQL_FIELD-facing function now does the JSON mariadb_field_attr check and the ENUM/SET flag check at the C-struct level, then delegates the type-code switch to mariaDBTypeName. Tests call mariaDBTypeName directly with primitives.

What is NOT covered by these tests (and why)

• fix(plugins): MySQL numeric column rendering and Query cancelled alert race #1209 Cancellation alert race — requires the query executor coordinator with a live driver to demonstrate. Catch sites now filter error is CancellationError || Task.isCancelled defensively; integration test only.
• perf(datagrid): scroll-path improvements cut max main-thread stall by 62% #1212 P0 feat: Query history #1 skip needsDisplay when unchanged + fix(datagrid): redraw cell when editability changes (chevron flicker on save/refresh) #1214 redraw on editability change — observable only through NSView.needsDisplay, which AppKit consumes asynchronously. Would need a snapshot or NSView mock harness.
• perf(datagrid): scroll-path improvements cut max main-thread stall by 62% #1212 P0 Centralize keyboard shortcuts following macOS HIG patterns #3 date parser memoization — pure performance optimization, no semantic change. Existing date-parse tests in the codebase already pin correctness.
• perf(datagrid): scroll-path improvements cut max main-thread stall by 62% #1212 P0 Fix data grid undo/redo and multi-row deletion bugs #4 autoreleasepool wrap + feat: implement tab persistence and last query memory #5 drop wantsLayer — implementation details verified via Instruments traces, no observable API change.
• fix(welcome): wire WindowOpener bridge from all scenes to fix missing-window race #1213 WindowOpener bridge wiring race — SwiftUI Scene lifecycle, can't fire .onAppear / .task in unit tests without a real App.
• fix(plugin-mysql): read MARIADB_CONST_STRING by length to stop JSON detection flicker #1216 MARIADB_CONST_STRING over-read — the fix is a buffer-bounded read of metadata that only comes from a live MariaDB connector. The over-read symptom depends on adjacent memory contents (allocator state); reproducing in a test would require either a real connection or fixture buffers planted with sentinel bytes.

Test plan

• xcodebuild test -only-testing:TableProTests/MariaDBTypeNameTests — all pass.
• xcodebuild test -only-testing:TableProTests/RowDisplayCacheTests — all pass.
• xcodebuild test -only-testing:TableProTests/PluginCellValueSortKeyTests — all pass (including new asText / asBytes cases).
• Full test suite still green: xcodebuild -project TablePro.xcodeproj -scheme TablePro test -skipPackagePluginValidation.
• Verified locally: ** TEST SUCCEEDED ** on the four target suites.