feat(timeseries): add time-series analysis extension with embedding similarity and drift detection

[chiangchenghsin-hash]

Summary

Adds a new timeseries extension with two table functions for time-series analysis on embedding vectors:

EMBEDDING_SIMILARITY

Compute cosine similarity and per-dimension feature similarity between two embedding vectors of any dimension (not hardcoded).

CALL embedding_similarity(
    CAST([0.1, 0.2, 0.3], 'DOUBLE[]'),
    CAST([0.1, 0.2, 0.4], 'DOUBLE[]')
) RETURN cosine_similarity, feature_similarity, dimension_count;

Returns: cosine_similarity ([-1, 1]), feature_similarity ([0, 1]), dimension_count.

DETECT_DRIFT_POINTS

Detect drift points in a sequence of embeddings by computing pairwise cosine distances between consecutive vectors. Generic labels (commit hashes, timestamps, version numbers) replace hardcoded chapter numbers.

CALL detect_drift_points(
    CAST([1.0, 0.0, 0.0, 0.0, 1.0, 0.0], 'DOUBLE[]'),
    2,
    CAST([100, 200], 'INT64[]'),
    3,
    threshold := 0.1
) RETURN label, drift_magnitude, significance, direction;

Returns: label (INT64), drift_magnitude (DOUBLE), significance (DOUBLE), direction ("up"/"down").

Use Cases

• Architecture drift: embeddings of code snapshots per commit → detect when architecture changed significantly
• Content drift: embeddings of document revisions over time → detect when content drifted
• Behavior drift: embeddings of API response patterns → detect behavioral changes
• Vector comparison: compare any two embeddings regardless of dimension

Design

• Pure C++ implementation, no external dependencies (no Arrow, no OpenMP)
• Uses LadybugDB's SimpleTableFunc pattern
• inferInputTypes callback for ANY → LIST<DOUBLE> type resolution
• Data prepared via Cypher COLLECT at app layer — no table scans needed
• All algorithms are deterministic (no randomness, no LLM calls)

Files

File	Purpose
timeseries/CMakeLists.txt	Top-level CMake
timeseries/src/function/CMakeLists.txt	Function library
timeseries/src/main/CMakeLists.txt	Extension library
timeseries/src/include/function/timeseries_function.h	Function registration structs
timeseries/src/include/main/timeseries_extension.h	Extension header
timeseries/src/main/timeseries_extension.cpp	Extension loader
timeseries/src/function/embedding_similarity.cpp	EMBEDDING_SIMILARITY implementation
timeseries/src/function/detect_drift_points.cpp	DETECT_DRIFT_POINTS implementation
timeseries/test/test_files/timeseries.test	E2E tests (3 cases)
timeseries/README.md	API documentation

Motivation

This extension was developed for architecture governance (detecting when code architecture drifts from the intended design via embedding comparison). It generalizes domain-specific bitemporal analysis functions into reusable primitives that work with any embedding space.