Improve validation and resource handling

Add Qdrant retry logic

Add percentile comment

Replace manual dict construction

Rename citation fields, strengthen test assertions from loose bounds

Add multi-failure test for evidence quality gate

Simplified cold-start, API takes query directly

Use canonical is_refusal() from faithfulness module

Replace custom percentile with np.percentile

Add integrity verification to cache thread safety test

Use MagicMock directly as default_factory for simple cases

Extract magic numbers to constants incl other polishing

Standardize metrics API to use seconds consistently

Harden middleware shutdown handling and security headers

Fix graceful shutdown resource handling

Harden summary.py

Calibrate evidence quality gate thresholds

Improve metric reporting

Add mixed evidence handling to explanation prompts

Add bootstrap confidence intervals to evaluation metrics

Refactor sanity checks with improved adversarial detection

Replace direct file writes to *_latest.json with atomic symlink pattern

Add retry logic for transient LLM failures

Add query sanitization for injection defense

Split CI into quick and full modes

Track rated_count incrementally

Add test coverage for shutdown, lifespan, and cache paths

Add partial result indicator and explanation timeout metrics

Expand /recommend/stream test coverage

Fix LLM retry timing metrics

Test POST /recommend/stream SSE event flow

Consolidate citation verification

Local make ci now matches GitHub CI

Clean up comments

Add structured logging

Add concurrent load testing for SemanticCache

Add singleton pattern to Qdrant client

Add RefusalType enum and query sanitization

Add punctuation normalization

Extract Qdrant pagination helper

Add boundary test for evidence quality gate

Improve low-relevance refusal message

Remove unused verify_explanation_full

Fix docstring

Fix precision_at_k

Normalize provider string input

Add empty list validation

Add explicit pydantic dependency

Reorder index creation before upload

Fix upload_chunks single-pass batching