Bug fixes

[bdurand]

Fixed

• sync_table_data! with delete_missing: true now raises an ArgumentError instead of deleting every row in the table when the data files contain no rows (for example, when a data file was accidentally emptied or truncated).
• Generated predicate methods (e.g. record.active?) now cast the data file value to the attribute type before comparing. Previously the raw file value was compared to the cast database attribute, so predicates silently returned false whenever the types differed (guaranteed for CSV data files, where all values are strings).
• Single table inheritance subclasses no longer raise NoMethodError from instance_names, instance_keys, protected_instance?, and other class methods. Subclasses now share the support table state defined on their base class.
• Named instance helper methods are now redefined when a later data file overrides an attribute or key value, so the helpers always return the merged values that are synced to the database. Previously they permanently returned the values from the first file that defined the named instance.
• named_instance_attribute_helpers can now be called again with an attribute that was already registered without raising an ArgumentError.
• YAML data files can now use anchors/aliases and date/time values. Previously these raised Psych::AliasesNotEnabled or Psych::DisallowedClass errors.
• protected_instance? no longer returns stale results when data files are added after the protected keys were first computed.
• Fixed broken cycle detection in the autosave association check during syncs that could cause infinite recursion on cyclic autosave associations.
• sync_table_data! now retries once on ActiveRecord::RecordNotUnique errors caused by concurrent syncs inserting the same rows from another process.
• sync_table_data! now returns an empty array instead of nil when the table does not exist.
• named_instance now raises a clear ActiveRecord::RecordNotFound error for undefined named instances instead of querying the database for a nil key (which could silently return a row with a NULL key value).
• Memoized class-level state is now consistently synchronized with the class mutex to avoid races on non-MRI Ruby implementations.
• Setting config.support_table.auto_sync = false before the gem is loaded is no longer overwritten back to true by the Railtie.
• The documentation tasks no longer corrupt model source files that contain duplicated generated YARD doc blocks (e.g. from a bad merge); the regex that finds the generated block is no longer greedy.
• Data file names containing extra dots no longer break the class name detection used by SupportTableData.sync_all! to eager load models.
• Error messages for invalid named instance definitions now include the model class name instead of repeating the instance name.