NEWS

jmotif 1.3.2 (2026-07-23)

Fixed: RRA phase-2 visit_array buffer sized to grammar interval count (was series length); could overflow and segfault when seed >= 0 enabled the seeded random-search path. Tier-B tests now exercise seed = 0 on ecg0606.
Tests: tier-B RRA conformance pins on ecg0606 — exact span and NN distance at w100 and w120 (seed = 0); w150/p7/a4 asserts the top discord is not a zero-distance boundary gap.
Docs: README lists CRAN install first; fix header anchors for SAX-VSM and RePair; rename RePair walkthrough to §7.1; link saxpy 2.0.1 on PyPI; roxygen \examples{} on sax_via_window, SAX-VSM bag/TF·IDF helpers, and cosine_sim; clarify cosine_dist() title and find_discords_rra() defaults in .Rd files; Codecov badge URL → app.codecov.io (CRAN incoming check).
Packaging: ship NEWS.md in the tarball; exclude agent audit HTML from builds; refresh DESCRIPTION summary; roxygen2 8.0.0 metadata.

Memory and performance (Phases 1–4):
- Fix memory leaks in RePair grammar (rule_record by-value), RRA, and cosine_sim (Valgrind-clean on long runs).
- HOT-SAX: precompute z-normed sliding windows (~1.4× on ecg0606, ~17× on longer series); identical discord positions and distances.
- SAX sliding window: _paa2 in-place fractional PAA, _znorm_slice, unified _sax_via_window path (~2.3× on bag construction).
- RRA: fused normalized distance, precomputed w_size z-norm windows (~1.3× on ecg0606); identical results and distance-call counts.
Safety: idx_to_letter() validates index range [1, 26].
Internal: use _is_equal_mindist(std::string) in SAX numerosity reduction; remove unused HOT-SAX state; clarify cosine_dist() roxygen.

Matrix Profile / MASS convention:

znorm now uses POPULATION standard deviation (divide by n) instead of sample std (n-1). Each normalized window then has empirical variance exactly 1, the assumption behind SAX's equiprobable Gaussian breakpoints.
SAX symbol assignment maps a value exactly on a breakpoint to the symbol ABOVE the cut (cuts[j] <= ts[i]), matching the jmotif Java original and saxpy. Previously a strict < mapped on-cut values to the symbol below.
PAA: fixed a floating-point boundary bug in the fractional PAA (_paa2, the public paa() / sax_by_chunking path). The final segment break paa_num * (len/paa_num) can round above len in IEEE-754 (e.g. 7*(29/7) = 29.000000000000004), which dropped the last sample from the final segment and corrupted its mean on ~5-6% of non-divisible (n, paa_num) shapes. The final break is now snapped to len, matching the saxpy / Java fractional PAA. NOTE: all three are behavioral changes -- SAX words for windows whose normalized PAA values fall near a breakpoint may differ from 1.2.x output.

HOT-SAX and brute-force discord discovery now measure the nearest-neighbour distance on Z-NORMED windows (the SAX premise of shape similarity):
- find_discords_hotsax: the three internal _znorm() calls had been left commented out, so distances were computed on RAW subsequences -- it reported raw-amplitude outliers (e.g. ecg0606 pos 411 / 1.50) instead of true z-normed shape discords (pos 430 / 5.28). The windows are now z-normed and compared with an early-abandoning Euclidean distance.
- find_discords_brute_force now z-normalizes too and gained an n_threshold parameter (default 0.01); previously it had no threshold and ran on raw subsequences. It is again a valid z-normed reference for HOT-SAX.
- Both now use a deterministic lowest-index tie-break and a +/-(w-1) exclusion zone, so results are reproducible and match saxpy / Java exactly (ecg0606 top discords: 430/5.279, 318/4.176, 2080/2.393). NOTE: behavioral change -- discord positions/distances differ from 1.2.x, which reported raw-distance discords.

Pre-release development version.
Added the Re-Pair grammatical inference algorithm implementation.
Added grammatical "rule coverage" curve implementation facilitating a time series anomaly discovery.
Added RRA (Rare Rule Anomaly) algorithm implementation for grammar-based variable length time series anomaly discovery.

Added the HOT-SAX time series discord finding algorithm implementation.
Added the brute-force time series discord finding algorithm implementation based on the early-abandoning Euclidean distance.