…taSQL

getDown() previously called getOldValue() unconditionally, which crashes
with a fatal error when the diff contains DiffOpAdd objects (which lack
that method). Also, getDown() did not null-guard getOldValue(), causing
NULL column values to be written as empty strings instead of SQL NULL.

Both getUp() and getDown() now:
- Check for DiffOpAdd/DiffOpRemove to avoid calling missing methods
- Null-guard return values to emit col = NULL instead of col = ''

Added comprehensive unit tests covering DiffOpChange, DiffOpAdd,
DiffOpRemove, null values, and mixed DiffOp types in a single UPDATE.

Related contributor PR: #93

MySQL data diff queries in LocalTableData now use backtick-quoted
`db`.`table` references instead of bare db.table. This prevents
MySQL from interpreting hyphens in database names (e.g. 'my-db') as
subtraction operators, which caused PDO syntax error 1064.

Fixed in getOldNewDiff() (2 queries) and getChangeDiff() (1 query +
key column references).

Added unit test verifying MySQLDialect::quote() correctly handles
hyphenated names, names with backticks, and names with spaces.

Related contributor PR: #92

TableIterator::next() now calls ->toArray() on the Illuminate Collection
returned by ->get(), and normalises stdClass rows to associative arrays.
This prevents array_merge() warnings and silent data loss in ArrayDiff
when the Collection was incorrectly passed as-is.

Added unit tests for ArrayDiff verifying:
- Identical data produces no diff
- New rows are detected as DiffOpAdd
- Deleted rows are detected as DiffOpRemove
- Changed rows are detected with correct diff keys

AlterTableEngineSQL::getUp() and getDown() now return empty string when
the engine or prevEngine value is empty/null, preventing invalid SQL
like 'ALTER TABLE `t` ENGINE = ;'.

Also hardened TableSchema::getDiff() to skip engine diff creation when
either engine value is empty (e.g. views that have no ENGINE).

Added unit tests for valid engines, null engines, and empty string engines.

MySQLAdapter::getTables() now uses SHOW FULL TABLES WHERE Table_type =
'BASE TABLE' instead of SHOW TABLES, preventing views from entering the
diff pipeline and being incorrectly handled as tables (which would
generate DROP TABLE instead of DROP VIEW).

PostgreSQL and SQLite adapters were already correct — pg_tables only
includes base tables and SQLite filters type = 'table'.

Added unit tests for DropTableSQL and AddTableSQL output verification.

Related contributor PR: #123

MySQL CONCAT() returns NULL if any argument is NULL, causing SHA2(CONCAT(...))
to produce NULL for any row with nullable columns. Two genuinely different
rows both hash to NULL and compare as equal — silently dropping real
data differences from the diff output.

Fix: wrap each column in IFNULL(col, '\0') inside the CAST before
concatenation, and add a NULL-presence bitmap (IF(col IS NULL, '1', '0'))
as a secondary comparison column. The WHERE clause now checks both
hash1 <> hash2 OR nullmap1 <> nullmap2, so:
- NULL vs empty string are distinguished (different bitmaps)
- NULL vs non-NULL values are caught (different hashes + bitmaps)
- Regular value changes continue to work via hash comparison

Related contributor PRs: #77, #63

InsertDataSQL::getUp() and DeleteDataSQL::getDown() now emit explicit
column name lists: INSERT INTO `t` (`col1`,`col2`) VALUES(...)
instead of positional INSERT INTO `t` VALUES(...).

This prevents silent data corruption when source and target tables have
identical columns in different order (common after ALTER TABLE ADD on
different servers). Also handles NULL values correctly in the column list.

Added unit tests covering:
- Column names included in INSERT output
- NULL value handling in INSERT
- DELETE UP/DOWN direction correctness

Remove both ini_set('memory_limit', '512M') calls from DBDiff::run()
and DBDiff::getDiffResult(). The memory limit should be controlled by
the user's php.ini or CLI -d flag, not overridden by the library.

Add MemoryLimitTest to verify the source no longer contains
ini_set('memory_limit', ...) and that autoloading DBDiff does not
change the runtime memory_limit.

Add a Fix Status summary table at the top of bugs.md documenting:
- Commit hash for each bug fix
- Brief description of what was changed
- Test file references
- Note that bugs 2 and 9 were combined (same file/method)

Original bug detail table preserved below for reference.

Bug 8 changed INSERT statements from positional VALUES to explicit
column lists:
  Before: INSERT INTO "posts" VALUES('3','2',...)
  After:  INSERT INTO "posts" ("id","user_id",...) VALUES('3','2',...)

Update all affected PostgreSQL (14-18) comprehensive test baselines
and the end2end baselines for both PostgreSQL and SQLite to match
the new correct output. MySQL baselines were unaffected.

The CLI entry points (dbdiff, dbdiff.php / PHAR) now set a default PHP
memory limit of 1G on startup. PHP's built-in default of 128M is too low
for diffing realistic database sizes.

The limit is fully configurable at three levels (highest wins):
  1. --memory-limit=<value>  CLI flag (e.g. --memory-limit=2G)
  2. memory_limit: <value>   top-level key in .dbdiff / dbdiff.yml
  3. 1G default in the entry point scripts

Any PHP shorthand is accepted (512M, 1G, 2G, -1 for unlimited).

The ini_set is placed in the CLI entry points only — not in the library —
so programmatic consumers (e.g. embedding DBDiff as a Composer dependency)
are never affected.

For native binary users (@dbdiff/cli npm package) ini_set has no effect,
and they are unaffected.

Bug 8 changed InsertDataSQL::getUp() and DeleteDataSQL::getDown() to emit
explicit column-name lists. Update all 7 SQLite comprehensive test baselines
to match the new format.

Also fix createTestConfig() to create tests/config/ if absent, and add a
.gitkeep so the directory is tracked in git. Without the directory the config-
file-based test cases silently skipped, leaving their baselines stale.