Phase 2: Universe validation (check_universe_appropriateness with
coverage/many-to-many diagnostics), user set redefinition
(redefine_user_sets matching R LOLA's redefineUserSets), and
restricted universe construction (build_restricted_universe with
interval merging).

Phase 3: RegionDB struct wrapping Igd + original region sets +
annotations. Loads LOLA-format folder structure (collection.txt,
index.txt, regions/*.bed), builds IGD in one pass. Supports
collection filtering, file limits, merge, and BEDbase integration
via from_igd_with_regions.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Phase 4: Rust extendr bindings (lola.rs) exposing 7 functions -
loadRegionDB, loadRegionDBFromBeds, runLOLA, listRegionSets,
checkUniverseAppropriateness, redefineUserSets, buildRestrictedUniverse.

R wrappers (lola.R) match original LOLA package signatures for
drop-in compatibility. Accepts RegionSet S4 objects, returns
data.frames with LOLA result columns.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Phase 5: PyO3 bindings as feature-gated lola submodule
(from gtars.lola import RegionDB, run_lola). RegionDB class
with from_folder/from_bed_files constructors. Regions passed
as native Python tuples for ergonomic API. Returns column-oriented
dicts for easy DataFrame conversion.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Implements Benjamini-Hochberg FDR correction per user set, TSV output
matching R LOLA format, and adds qValue to Python and R bindings.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- Load all BED files from regions/ directory, not just index.txt entries
- Use binary per-region overlap counting (count_region_hits) to match
  R LOLA's countOverlaps semantics
- Use survival function sf() instead of 1-cdf() to avoid catastrophic
  cancellation in p-value computation for highly significant results
- Fix R wrapper symbol names and regenerate extendr/roxygen docs
- Add compare_lola.Rmd test script

All contingency table values now match R LOLA exactly. P-values match
to within 2.4e-11 (machine epsilon).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Phase 3 of interval operations: IGD-backed overlap queries on RegionSet.

- Add Igd::from_single_region_set, find_overlaps_regionset, and
  count_overlaps_per_query to gtars-igd with deduplication across tiles
- Add find_overlaps and count_overlaps to IntervalRanges trait
  (gtars-genomicdist now depends on gtars-igd)
- Add R FFI wrappers and S4 methods importing generics from IRanges
- Refactor gtars-lola universe.rs to use RegionSet methods instead of
  hand-rolled overlap code (~80 lines removed)
- All 245 Rust tests and 7/7 LOLA comparison checks pass

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Expose regiondb_anno(), regiondb_collection_anno(), and
regiondb_region_set() so R consumers can access annotations and
individual region sets from a gtars-loaded RegionDB without needing
to re-read from disk. This enables episcope and other packages to
use gtars as a drop-in backend while retaining access to metadata
needed by post-enrichment functions like extractEnrichmentOverlaps.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

# Conflicts:
#	gtars-genomicdist/src/interval_ranges.rs
#	gtars-python/Cargo.toml

…tions

- gaps(): add leading gap from position 0 to first region per chromosome,
  matching R's gaps() behavior (Fix 1)
- Add CoordinateMode enum (Bed/GRanges) in gtars-core for coordinate
  convention differences between R (1-based closed) and BED (0-based half-open)
- Region::mid_point_with_mode() uses banker's rounding for GRanges mode (Fix 3)
- calc_summary_signal() formats labels with 1-based start in GRanges mode (Fix 2)
- Thread CoordinateMode through TssIndex, calc_tss_distances,
  calc_feature_distances; R bindings use GRanges, all others default to Bed

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- calc_neighbor_distances: sort regions by (start, end) per chromosome
  before windowing; filter out non-positive distances (Fix 4)
- calc_nearest_neighbors: sort regions per chromosome; skip single-region
  chromosomes instead of pushing u32::MAX sentinels (Fix 5)
- Update tests to reflect new behavior

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- Add region_distribution_with_chrom_sizes() that creates bins based on
  each chromosome's actual size instead of uniform global bin size (Fix 6)
- R binding accepts optional chrom_names/chrom_lengths parameters
- Tourney dispatch passes chromSizes from getChromSizes("hg19") to gtars
- R distribution() generic gains optional chromSizes parameter

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

calc_expected_partitions now computes partition genome fractions using
priority resolution: each genomic bp is assigned to the first partition
that covers it (via reduce + setdiff), preventing double-counting when
partitions overlap. This matches R's approach. (Fix 7)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- Fall back to collection-level description when file-level description
  is empty, matching R LOLA's behavior (Fix 8)
- Add 'size' column (region count per file) to regionDBAnno output

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…sums

Three bugs in calcExpectedPartitions that caused expected value mismatches:

1. Strand-aware setdiff for proximal promoters: into_regionset() was called
   before setdiff(), dropping strand info. R's setdiff is strand-aware, so
   +/- strand promoters at the same position were incorrectly subtracted.

2. trim() silently dropped regions on chromosomes not in chrom_sizes (e.g.,
   chrMT vs chrM naming mismatch). Now keeps such regions unchanged, matching
   R's trim() behavior.

3. R binding pre-reduced genes before computing promoters, collapsing
   overlapping genes and losing their individual promoters. R computes
   promoters(rawGenes) first, then reduces. Removed the pre-reduce.

Also adds promoters_stranded() to preserve strand through the promoter
pipeline, documents the R neighbordt() bug in an Rmd, and updates the
dispatch to pass chromSizes for calcExpectedPartitions.

Results: 5/8 tourney functions exact match, 3 intentional divergences
(2 where gtars is more correct than R, 1 deliberate design choice).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- IGD: 3 new tests comparing old vs new API (single-file, multi-file,
  disk format compatibility)
- LOLA: 10 new tests for enrichment edge cases, FDR correction, and
  database loading
- gtars-py: disambiguate count_overlaps/find_overlaps where both
  IntervalRanges and RegionSetOverlaps traits provide them

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

… RegionSetOverlaps

- Remove count_overlaps/find_overlaps from IntervalRanges trait (genomicdist)
- Remove gtars-igd dependency from gtars-genomicdist
- Add min_overlap: Option<i32> to all 4 RegionSetOverlaps methods (overlaprs)
- Rewrite universe.rs to accept pre-built &Igd instead of rebuilding per call
- Switch R bindings to RegionSetOverlaps with min_overlap for countOverlaps/findOverlaps
- Update Python, WASM bindings for new signatures
- Fix loadRegionDB description fallback to use collection folder name (matches R)

Resolves E0034 trait ambiguity in gtars-py. Tourney: 24/27 pass.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…ings

- Fix iter_chroms() to preserve insertion order instead of HashSet
  non-deterministic order (fixes GC content correlation bug)
- Fix S4 dispatch: register methods for character/data.frame instead of
  ANY for Bioconductor generics (narrow, shift, etc.) to avoid hijacking
  GRanges dispatch
- Switch LOLA odds ratio from sample OR (a*d)/(b*c) to Fisher conditional
  MLE matching R's fisher.test()$estimate, using Brent root-finding on the
  noncentral hypergeometric distribution
- Add missing IntervalRanges wasm bindings (shift, flank, resize, narrow,
  disjoin, gaps, intersect) to match R bindings
- Update compare_lola.Rmd for LOLACore compatibility and oddsRatio check
- Update tutorial_regionset.Rmd with tabset comparisons and known diffs
- Gitignore LOLACore database and ext/ test data
- Remove test_dropin.Rmd
- Regenerate roxygen docs via rextendr::document()

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Use Igd::from_named_region_sets (takes &RegionSet references) instead of
from_region_sets (takes copied (String, i32, i32) tuples), avoiding a
full allocation pass over all regions during database loading.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Add shift, flank, resize, narrow, disjoin, gaps, and intersect
subcommands to match R binding coverage of IntervalRanges methods.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…API safety

Batch 1 (correctness):
- LOLA: Replace saturating_sub with validated subtraction + NegativeContingency error
- IGD: Guard from_region_sets/from_named_region_sets against end < start intervals
- LOLA: Use from_named_region_sets in merge() to avoid tuple copy overhead
- LOLA: Warn on skipped BED files instead of silent continue

Batch 2 (API safety):
- IGD: Guard add() against negative coordinates
- IGD: Clamp negative query coords in count_overlaps()
- IGD: Remove chr-prefix filter from parse_bed_line (support non-UCSC assemblies)
- IGD: Remove 321M coordinate cap from from_bed_dir

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…c comments

Batch 3 (code quality):
- IGD: Merge validation into parse pass, eliminating double file read
- LOLA: Remove unused EmptyUserSet error variant
- IGD: Fix TSV format (remove extra spaces, write avg_region_width as f64)
- genomicdist: Document bp-mode double-counting behavior in partitions
- IGD: Add # Panics doc comments for add(), save(), count_overlaps(),
  find_overlaps_regionset(), count_overlaps_per_query()

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…docs

- Fix resize("center") midpoint overflow and chromosome_statistics width
  sum overflow (u32→u64 accumulator)
- Fix boxplot_stats NaN panic and narrow() underflow with saturating_sub
- Replace disjoin() O(n²) coverage check with O(n log n) sweep-line
- Fix NaN odds ratio ranking (worst rank instead of arbitrary placement)
- Add FDR correction to Python and R LOLA bindings
- Fix R LOLA contingency values (i32→f64) to avoid truncation
- Fix R io.rs unwrap → proper error propagation
- Fix R genomePartitionList: cast f64 coordinates to integer for stranded path
- Revert R output types for counts/widths/distances back to i32 (only
  coordinates need f64 for u32 range)
- Fix R r_narrow type error propagation
- Remove dead sample_odds_ratio, leftover println
- Add IGD hits debug_assert, contigs pub(crate) with accessor
- Document IGD i32 coordinate limit and uniwig sorted-input requirement
- Add regression tests for overflow, NaN, depletion direction
- Add Python LOLA binding tests

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Switch CLI handlers and R bindings from legacy create_igd_f/igd_search
to the new unified Igd struct. Add Igd::from_bed_files() constructor
that accepts explicit file paths, enabling directory, .txt filelist,
and stdin input modes. R search now returns structured data (named list)
instead of TSV strings. Remove dead CLI search flags that were never
wired up. Legacy create.rs/search.rs kept in place.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Validates the unified Igd struct against legacy igd_t/igd_t_from_disk
using real genomic data (6 BED files: CpG islands, lamin B1 LADs, VISTA
enhancers across 2 collections) for broader coverage before merge.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>