- Create base OutputGenerator abstract class with GeneratorArgs
- Implement filters module for job filtering based on deployment type,
  full/nightly runs, and platform exclusions
- Add comprehensive CircleCI generator with support for:
  - Multiple architectures (x64, aarch64)
  - Sanitizer builds (tsan, alubsan)
  - Replication version 2 tests
  - UI tests with RTA
  - Docker image creation
  - Resource sizing based on sanitizer and cluster mode
  - Special cases for chaos and replication_sync tests

This provides a clean separation between parsing, filtering, and output
generation, making it easier to maintain and extend.

Major refactoring to improve code organization and testability:

**Refactored modules:**
- filters.py: Removed hardcoded platform detection, introduced PlatformFlags
  and FilterCriteria dataclasses for explicit configuration
- base.py: Split GeneratorArgs into focused dataclasses (TestExecutionConfig,
  CircleCIConfig, GeneratorConfig) following single responsibility principle
- circleci.py: Restructured into smaller focused methods, removed sys.path
  hacks, now uses proper relative imports
- sizing.py: New module extracting resource sizing logic with sanitizer
  overhead calculations

**Improvements:**
- Fixed all imports to use proper relative imports (..module pattern)
- Removed mutable default arguments with field(default_factory)
- Better separation of concerns across modules
- All code formatted with black

**Testing:**
- Added 32 new tests for filters.py (16 tests) and sizing.py (16 tests)
- All 111 tests passing (79 original + 32 new)
- Tests cover platform flags, filter criteria, deployment type filtering,
  gtest filtering, resource class mappings, and sanitizer overhead logic

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

**Refactoring:**
- sizing.py: Use ResourceSize enum instead of hardcoded strings
- sizing.py: Add architecture alias support (amd64, x86_64, arm64)
- sizing.py: Extract _apply_sanitizer_overhead() for better separation
- filters.py: Extract is_gtest_suite() and matches_deployment_filter()
- filters.py: Add GTEST_PREFIX constant, use field(default_factory)
- filters.py: Merge nested if statement

**Tests:**
- Update all tests to use ResourceSize enum
- Add test coverage for architecture aliases

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

Add dependency injection for environment and date access, support pre-loaded config dict, fix hardcoded size strings to use ResourceSize enum, and add 25 unit tests.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

Create generate_config_new.py that uses the refactored architecture to generate CircleCI configuration from test definitions. Supports all key features: test filtering, sanitizers, UI tests, and multi-file definitions.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

- Add init_driver_repo_command field
- Preserve extraArgs: prefix in argument keys
- Add optionsJson for multi-suite jobs with nested structure
- Handle moreArgv and suffix support
- Support duplicate jobs with unique keys
- Add filtering for full flag at job and suite level
- Fix workflow naming (with_ui_tests)
- Add non-maintainer build job for x64 enterprise
- Update tests for unique job keys

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

Old generator defaulted to medium for cluster/mixed and small for
single. New generator was always defaulting to small, causing size
mismatches for cluster tests without explicit size specification.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

- Add init_command field to RepositoryConfig for driver init scripts
- Support 'alubsan' as sanitizer alias (treated as 'asan')
- Fix comparison script to continue testing all configs even on failures

All test configurations now generate successfully with only cosmetic
differences (argument ordering) compared to old generator.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

Updated generate_config_new.py to only accept 'tsan' and 'alubsan' as
valid sanitizers, matching the old generator's behavior. Individual
sanitizers (asan, ubsan, lsan) are not accepted as they are not used
in production.

alubsan = asan + lsan + ubsan combined
Internally normalized to 'asan' for sizing logic.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

Added Sanitizer enum (TSAN, ALUBSAN) to replace string literals:
- Sanitizer.from_string() validates and parses CLI input
- BuildConfig.sanitizer now uses Optional[Sanitizer]
- FilterCriteria.full now uses Optional[Sanitizer] instead of bool
- Updated all code to use sanitizer.value when converting to strings
- Removed test for invalid sanitizer strings (now handled by enum)
- Updated tests to use Sanitizer.TSAN and Sanitizer.ALUBSAN

This provides type safety and prevents typos in sanitizer names.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

Added Architecture enum (X64, AARCH64) to replace string literals:
- Architecture.from_string() validates and parses input with alias support
  (x64, x86_64, amd64 → X64; aarch64, arm64 → AARCH64)
- BuildConfig.architecture now uses Architecture enum
- Updated CircleCI generator to use Architecture.X64 and Architecture.AARCH64
- Updated ResourceSizer to accept Architecture enum or string
- Updated all code to use architecture.value when converting to strings
- Updated tests to use Architecture.X64 and Architecture.AARCH64

This provides type safety and prevents typos in architecture names.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

- Separated FilterCriteria.full (bool) from sanitizer (enum) - they are independent concepts
- Fixed workflow generation to use correct sanitizer field
- Added container_suffix and init_command support for driver tests
- Fixed job-level arangosh_args parsing
- Added timeLimit for nightly chaos tests
- Extracted _generate_unique_key() helper to reduce duplication
- Added remove_extra_args() to normalize script for cosmetic diffs

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

After merge with devel:
- Added job_type field to TestJob for custom CircleCI job templates
- Parse 'job' field from test definitions (defaults to 'run-linux-tests')
- Updated test file paths (arangojs-test-definitions.yml → arangojs.yml, etc.)
- Fixed FilterCriteria to properly separate full and sanitizer fields
- All tests passing (139/139)

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

Adapted new generator to match changes from merge commit ec7f7ef:

Repository/Driver Config:
- Changed init_driver_repo_command to init_command to match old generator
- Only add init_command if field exists (even if empty string)
- Don't add empty driver fields when there's no repository config

ARM64 Workflow Generation:
- Removed sanitizer restriction for ARM64 workflows (now generates for all sanitizer configs)
- This matches the removal of the condition in the old generator

Argument Ordering:
- Fixed extraArgs order: optionsJson must come BEFORE replicationVersion and skipNightly
- This matches old generator behavior where optionsJson was added during parsing

UI Test Configuration:
- Changed rta-branch from empty string to None for proper YAML null serialization
- Added trailing space to UI filter strings to match old generator format

These changes bring the new generator output into alignment with the old generator
after the recent merge from devel branch.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

The old generator was deleting test["arangosh_args"] after first use,
which caused these settings to be lost when generating the aarch64 workflow.
This meant that memory/GC settings specified in test definitions only applied
to x64 builds, not aarch64 builds.

Removed the deletion to ensure arangosh_args are preserved and applied
consistently across all architectures.

This brings the old generator into alignment with the new generator's
correct behavior, eliminating the remaining test failures.

All 5 comparison tests now pass.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

- Extract generic enum parsing helper to eliminate duplication across 4 enums
- Add merge_field helper to TestOptions.merge_with() to reduce repetition
- Move parse_args_string to TestArguments class for better cohesion
- Consolidate CircleCI special case overrides into documented section
- Fix missing fields (suffix, full, coverage) in TestOptions.merge_with()

All 139 tests pass.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

Add support for overriding git branches in driver test definitions via CLI.
Example: --test-branches=go=feature-branch:js=main

When a test definition filename matches a prefix in test_branches (e.g.,
"go.yml" matches "go"), the specified branch overrides the repository's
default branch for all jobs in that file.

This completes feature parity with the old generator as specified in the
implementation plan (line 108).

All 139 tests pass.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

- Add time_limit field to TestOptions
- Parse timeLimit from YAML options (using camelCase key)
- Remove hardcoded _get_time_limit_override method
- Use job.options.time_limit directly in generator

This allows any job to specify a custom timeLimit in the YAML,
not just chaos. The chaos job already has timeLimit: 9600 in
test-definitions.yml (line 703).

All comparison tests pass.

When a test definition filename doesn't contain a directory
separator, automatically prefix it with "tests/". This matches
the behavior of the old generator and allows users to specify
just the filename instead of the full path.

- Add TypeVar for generic enum helper function
- Add type annotations for dict variables to fix mypy inference issues
- Assert buckets type in get_bucket_count to satisfy mypy
- Replace string architecture values with Architecture enum in test_sizing.py
- Install types-PyYAML for proper type checking

Remove backwards compatibility for string architecture values. The method now
only accepts Architecture enum, which provides proper type safety. This change
removes the ARCH_ALIASES dict and _normalize_arch() helper method that were
defeating the purpose of using enums.

Tests updated to exclusively use Architecture enum values instead of strings.

- Fix _dict_to_options_json to strip 'extraArgs:' prefix
  - Keys like 'extraArgs:log.level' now produce {'log.level': value}
  - Previously created {'extraArgs': {'log.level': value}}
- Fix job creation to skip jobs when all suites are filtered out
  - _create_test_job now returns None when filtered_suites is empty
  - Caller filters out None values to avoid invalid job configurations

- Reduce test_parsing_integration.py from 280 to 80 lines (71% reduction)
- Remove redundant per-file parsing tests (covered by test_all_yaml_files_parse_successfully)
- Remove trivial property tests (not integration test concerns)
- Remove filtering tests (belong in unit tests)
- Keep essential integration tests:
  - All files parse successfully
  - Main test structure validation
  - Multi-suite auto bucket behavior
  - Driver repository configuration
- Fix pylint errors in generate_config.py:
  - Move dataclasses.replace import to top level
  - Add pylint disable for Click decorator behavior
- Tests reduced from 147 to 132 (redundant tests removed)
- All remaining tests pass

Consolidate repetitive test patterns across unit tests using
pytest.mark.parametrize to reduce duplication while maintaining coverage.

Changes:
- test_filters.py: Consolidate deployment type and suite filtering tests
- test_sizing.py: Parametrize architecture mapping tests
- test_config_lib.py: Consolidate enum conversion and validation tests
- test_parsing_integration.py: Add MAIN_TEST_DEFINITIONS constant

Impact:
- Reduced total lines from 1976 to 1484 (25% reduction)
- Test count remains 157 (no coverage loss)
- Improved maintainability and extensibility

Create conftest.py with reusable fixtures for common test objects:
- BuildConfig fixtures (x64/aarch64, enterprise/community, with/without sanitizers)
- mixed_suite_job fixture for filter testing
- generator_factory fixture for CircleCI generator creation

Changes:
- tests_unit/conftest.py: New file with 104 lines of shared fixtures
- test_filters.py: Use mixed_suite_job fixture (517 -> 487 lines, -30 lines)
- test_sizing.py: Use BuildConfig fixtures (297 -> 295 lines, -2 lines)

Impact:
- Centralizes test object creation
- Reduces duplication across test files
- Makes tests more readable and maintainable
- Easier to extend with new fixtures in the future

We no longer support community builds - only enterprise builds are
maintained. This commit removes all community/enterprise distinction:

- Remove enterprise field from BuildConfig dataclass
- Simplify CircleCI generator to always use enterprise edition
- Remove community-related test fixtures and tests
- Update all BuildConfig instantiations

All 153 tests pass.