[nuk3s]

This PR significantly improves robustness by adding comprehensive input validation at load time and better error handling throughout the codebase.

Summary

The changes add early validation of corpus, gold standard, and adversarial probe files to catch configuration errors before they cause runtime failures. HTTP adapter code is consolidated into a shared module to ensure consistent error reporting. The readiness polling is improved to sample across the entire corpus rather than just the head. Query retry logic is refined to avoid retrying client errors that will fail identically.

Key Changes

Input Validation

• Added load_corpus() and load_gold() functions that validate row structure at load time, reporting line numbers and specific missing fields
• Added validate_probe() for adversarial probes, checking corpus/checks/corrections structure and regex validity before the run starts
• Enhanced load_jsonl() to report line numbers on JSON decode errors
• All validation failures now exit before any writes to the store, preventing orphaned run logs

HTTP Adapter Consolidation

• Created new memeval/adapters/_http.py module with shared request_json() helper
• Refactored Hindsight, Mem0, Letta, Zep, and ExampleRest adapters to use the shared module
• Ensures HTTP error responses include the server's error body (up to 200 chars) in all adapters for better debugging

Runner Improvements

• Added _spread() function to sample corpus docs evenly across the entire range (first and last always included) rather than just the head, catching silent write losses at the tail
• Refined retry logic with _retryable() to distinguish between retryable errors (5xx, 408, 425, 429) and client errors (4xx) that fail identically
• Updated readiness poll to use the spread sampling strategy

CLI and Event Handling

• Updated cli.py to use new load_corpus() and load_gold() functions
• Fixed cmd_adapters() to handle adapters with empty docstrings without IndexError
• Added read_events() robustness to skip non-object lines in run logs (arrays, scalars, etc.)
• Added python -m memeval entry point via __main__.py

Test Coverage

• Added comprehensive tests for malformed input files (corpus, gold, probe)
• Added tests for HTTP error handling across all adapters
• Added tests for readiness polling with spread sampling
• Added tests for retry logic distinguishing retryable vs non-retryable HTTP errors
• Added tests for negation handling in adversarial checks
• Added tests for CLI robustness (missing files, empty docstrings, etc.)

Dependencies

• Updated Textual requirement from >=0.80 to >=2.0
• Added dev optional dependency group with pytest and pytest-asyncio
• Updated CI to test against Python 3.10-3.13

Notable Implementation Details

• Hindsight batch size now has a minimum of 1 to prevent range() errors
• Readiness polling now samples up to 10 docs spread evenly across the corpus, ensuring tail batches are validated
• HTTP error bodies are truncated to 200 chars to keep error messages readable
• Malformed answer regexes in gold files remain runtime errors (per README contract) rather than load-time rejections
• All validation errors include row numbers for easy debugging