v1.6.0

• HTOK ISO Date Ingestion (rec_axis_store_typed):
  • Exported rec_axis_store_typed allowing direct ingestion of ISO date strings and epoch intervals from libhyle-source's .tok binary containers into the interval index.
• Linker Purity & Portable Build:
  • Removed hardcoded /home/ and -Wl,-rpath linker flags from root and test Makefiles.

Verification

gpg --verify CHECKSUMS.asc
sha256sum -c CHECKSUMS

v1.4.0

• Cormr Recall Kernel Integration: Updated to consume <ttypt/corm.h> and <ttypt/cormr.h>, aligning with the domain-free Recall Kernel architecture.
• Lemma Text Indexing Alignment: Updated ecosystem references and test fixtures from libstoma to liblemma.
• Memory Footprint Optimization: Scaled down initial allocation footprints for interval and split buffers.

1.3.0

• Renamed libit → libjoint: the it_* API is now joint_* (include/ttypt/joint.h, it.pc → joint.pc; LIBIT_LIMITATIONS.md → JOINT_LIMITATIONS.md). Renames span the whole walker/splitter read model while the interval-tree semantics stay intact.
• Kernel composition: new rec_axis_* adapters (rec_axis_fill_interval, rec_axis_open, plus rec_axis_store/rec_axis_unstore/rec_axis_readback) plug joint's interval/overlap matching straight into the recall kernel with rec_ref_t (uint32_t) refs; all axis parameters are CLI parameters.
• Sweep-line performance tiers: public API unchanged; all work is in the query/split path (joint_iter → joint_next) — -O3 -mpopcnt -mavx2 -mfma build (was -O0), pointer-cast split compares (no 16-byte memcpy per comparison), a query-level match_arena replacing per-match mallocs, a block-list split_arena replacing per-split mallocs, and an ephemeral who_set (open addressing) replacing the per-gap temp corm. Query benchmarks: 1k overlap query 48.5×, 3k overlap query 48×, 1000 insert/query cycles 6.8×, sparse query 6.2×, splits(100) 5.5× faster (medians; make bench).
• Bug fixes: two latent signed-integer overflows in src/test_extended.c exposed by UBSan (int overflow in timestamp arithmetic, int64 overflow in query-end) — both corrected to time_t/huge arithmetic.
• libqmap → libcorm rename.

Performance (medians, µs; make bench on one loaded box)

Baseline = pristine -O0 build (medians of 3); optimized = this tree
(medians of 9). Same machine, same load window.

Bench	Baseline	Now	×
10k insert	35 734	16 137	2.2
query-all (10k)	3 673 044	1 601 764	2.3
1k overlap insert	9 234	5 258	1.8
1k overlap query	5 444	326	16.7
sequential 5k	2 267 215	1 287 717	1.8
sparse query	81	13	6.2
splits (100 ent)	343	62	5.5
1000 insert/query cycles	180 756 539	26 737 804	6.8
15k insert	45 691	21 613	2.1
15k range query	2 034 606	997 656	2.0
3k overlap insert	52 867	30 614	1.7
3k overlap query	62 079	1 281	48.5
20k insert	93 440	33 995	2.7
20k range query	1 683 350	801 620	2.1

Absolute numbers swing 2–5× with machine load (both columns measured in
the same window, so ratios are like-for-like). All 74 tests pass
(./test.sh); ASan+UBSan clean (manual build, detect_leaks=0 — the
iterator arena is intentionally never freed; no teardown API exists, as
before). Valgrind out of scope per project policy.

Verification

gpg --verify CHECKSUMS.asc
sha256sum -c CHECKSUMS