CHANGELOG.md

Changelog

Unreleased

0.6.3 (2026-07-26)

Added

• demux is now part of the default cli feature, so the standard escpod build ships barcode demultiplexing without a rebuild. It pulls in zero new third-party crates (ndarray/serde_json/rayon/uuid were already in the graph). The accelerator features train, gpu, cnn-detect, and cnn-gpu remain opt-in because they need extra toolchains or a CUDA runtime (#150).

Fixed

• mad_normalize no longer aborts on a constant signal. A dead pore or flat read produced a zero MAD that panicked; since release builds use panic = "abort", a single bad read killed a multi-hour job (#150).
• Short fingerprints are no longer emitted silently. extract_fingerprint_from_signal treated keep_last as a maximum rather than an exact width, so a truncated fingerprint could reach the classifier. GBM errored out, but DTW/SVM scored the truncated query and reported a confident wrong barcode. Observed in real data at 1 ragged row in 9,894 (#150).
• read_query_csv validates row width. Malformed cells were dropped silently; the labeled loader had guarded this for a while, the query loader never did (#150).
• A GBM classify failure no longer discards its whole chunk. One failing read dropped all 1,024 reads in the chunk instead of just itself (#150).
• filter/subset write real read IDs into the signal table. The read_id column was zero-filled, diverging from ONT's own tooling and from the schema's documented "UUID for consistency checking" purpose. Files still loaded and round-tripped losslessly because the reads table is the authority for the read→signal mapping, so nothing read the column — but any tool that used it for the check it is named for would have rejected every file escpod filter/subset ever produced (#151).
• demux train output is reproducible. Grouping iterated a HashMap, so std_dev summation order — and its low bits — varied run to run, and TrainingOutput::barcodes was a HashMap, so serde emitted barcodes in a different order each process. Results are now sorted by read index and the map is a BTreeMap; three consecutive runs are byte-identical (#152).

Performance

• Fused demux pipeline is 2.3× faster on GBM models (151.0 s → 65.1 s on 1.22M reads / 10.4 GB, 48 cores; CPU utilization 511% → 1273%), and 1.27× on DTW-SVM. The pipeline was neither compute- nor I/O-bound — it ran faster at 8 threads than at 48, because block fill never overlapped with processing and the dominant barcode's writer channel blocked rayon workers that could not then be stolen from. Per-barcode counts are unchanged, and the staged detect/fingerprint subcommands still produce byte-identical output (#150).
• DTW-SVM classify is faster, bit-identically. DTW_LANES 16→32 and decision_function now skips support vectors that cannot contribute to a class pair: per-read predict 150.1 → 110.4 µs (−26.4%), and the 20-class decision function 366.9 → 26.7 µs (−92.7%), which matters for wider barcode designs (#150).
• filter is 2.29× faster and subset 1.74× (51.3 s → 22.4 s and 37.9 s → 21.8 s on 1.22M reads / 10.4 GB, 48 CPUs). The signal-table copy ran inline with the IPC writes on one thread, so every page fault stalled the thread that then issued the write — 75% of profile samples in memmove plus kernel page-fault time at 24% CPU. Batches now build across rayon and are written in order, with lookahead bounded to 16 batches in flight. merge is deliberately untouched: it copies whole Arrow IPC blocks and patches footer offsets, so it never rebuilds batches (#151).
• demux train is 8.7× faster (231.5 s → 26.5 s on 200k assigned reads, 48 cores; CPU utilization 86% → 1199%). extract_fingerprints parallelized only across files, so the common single-POD5 training run was effectively serial regardless of -t. It now walks files sequentially — one ascending mmap sweep per Arrow batch, preserving the single-stream I/O — and parallelizes inside each batch. Consensus fingerprints are bit-identical (#152).
• demux train-svm no longer computes an all-pairs DTW distance matrix and its RBF kernel matrix. Both were discarded: their only consumer took the kernel as an ignored parameter, so ever since the SMO fit was removed the emitted model has been a function of the labels alone. At N=50,000 the command now runs in 0.27 s / 25 MB, where the matrices alone would have wanted ~40 GB — the memory wall that --max-per-class exists to work around (#152).

Changed

• The gpu DTW-classify feature is now marked experimental. On a full node it measured slower than the CPU: 113.0 s on 64 cores versus 132.4 s with --gpu on an A30 (0.85×), plus ~2.2 GB more RSS. The apparent win disappears once the CPU gets the whole node instead of 16 of 64 cores. GPU CNN adapter detection (--method cnn --gpu) remains the GPU path that pays off (#150).
• demux train-svm --window and --gpu now warn that they do not affect the fit, instead of silently appearing effective (#152).
• compute_distance_matrix, distance_to_kernel_matrix, and train_svm_from_distances stay public and tested, so a real SVM fit can be wired back in without a signature change (#152).

Build / Tooling

• The signal RecordBatch builder is shared between filter/subset and the incremental Writer (build_signal_batch + SignalRow in utils::table_builders) rather than duplicated (#151).
• New test_table_conformance guardrail in the CI compat suite snapshots all three embedded Arrow tables from a pod5-written and an escapepod-written file and diffs schema plus contents, resolving signal read_id through each read's signal_rows. Validated as a guardrail: it fails against a pre-fix binary. Adds an ESCPOD_BIN override so the suite can target an arbitrary build (#151).
• Dependency bumps: clap 4.6.4, libc 0.2.189, and actions/setup-python 7.

0.6.2 (2026-07-22)

Build / Tooling

• mold is now the linker full time inside pixi. The base [activation] block exports LD_PRELOAD=$CONDA_PREFIX/lib/mold/mold-wrapper.so (+ MOLD_PATH) for every environment, so any pixi run [-e <env>] cargo … (and maturin) links with mold — no more mold -run wrappers or -e dev-only mold. Works with the stock system gcc 11.5 on the compute nodes (no -fuse-ld=mold support needed) via mold's mold-wrapper.so interposer, and needs no glibc-static. CI release builds (musl, outside pixi) are unaffected.

Added

• Auto-warmed read-id index on Python context-manager entry. Entering a Reader or DatasetReader in a with block now builds the in-memory read-id index, so repeated reads(selection=…) lookups take an O(k) indexed path instead of re-scanning the reads table each call (~2× faster random-access selection, on par with the pod5 package). Plain open() is unchanged — the index stays lazy for single-pass streaming. Size-gated by read count (default 5,000,000, overridable via ESCAPEPOD_AUTOINDEX_MAX) to bound memory; this is the in-memory index, not the .p5i sidecar, so no sidecar file is written (#97).

• PyPI publishing for the escapepod Python package. The release.yml workflow now builds abi3 wheels (CPython 3.9+) for Linux (x86_64/aarch64, manylinux + musllinux) and macOS (x86_64/arm64) plus an sdist, and publishes them to PyPI via Trusted Publishing (OIDC) on each v* tag. pip install escapepod will work once the first tagged release lands. The bindings crate gained abi3-py39 and complete PyPI metadata (readme, license, URLs, classifiers, type stubs).

Performance

• Resquiggle hot paths (band construction and refinement) iterate with array_windows instead of manual index arithmetic, dropping redundant bounds checks from the inner loops (#61).

Fixed

• Status output no longer escapes ANSI. The tracing formatter was escaping ANSI control sequences, so styled status/progress output rendered as literal escape codes; it now emits proper styling (#142).
• Quiet SIGPIPE on piped output. Piping CLI output into a consumer that closes early (e.g. | head) no longer produces a broken-pipe error — the CLI exits cleanly on a closed downstream pipe (#140).

0.6.1 (2026-07-20)

Fixed

• POD5 archives are now written atomically: output is staged to a temporary file and renamed into place on completion, so an interrupted or failed write can no longer leave a truncated/partial .pod5 at the target path.

Build / Tooling

• Dependency bumps: noodles-bam 0.91, noodles-sam 0.86, noodles-bgzf 0.48, noodles-csi 0.57, sha2 0.11, and the grouped cargo minor/patch and GitHub Actions updates.

0.6.0 (2026-07-11)

Added

• Native GBM (gradient-boosted tree-ensemble) barcode classifier for demux classify / fused demux, alongside the existing DTW+SVM path. Fingerprints are scored directly by a distilled tree ensemble (no DTW), loaded from the same auto-detected model surface. Fused-pipeline support is included, so escpod demux --model gbm.json … routes reads end-to-end.
• Architecture-agnostic GPU CNN/TCN adapter detection (demux detect --method cnn --gpu, feature cnn-gpu): runs any [B,1,L] -> [B,2,L] ONNX graph batched through onnxruntime's CUDA execution provider via the ort crate (load-dynamic — needs a CUDA-enabled libonnxruntime on ORT_DYLIB_PATH and a visible GPU at run time, nothing at build time). TCN detect is inference-bound, so this pays off (~7.6× end-to-end on an A30 at 20k reads). This is the ONNX-generic backend anticipated by the removal of the old arch-locked CUDA kernel below — it replaces it without hardcoding a topology.
• pod5-parity Python API: a multi-file DatasetReader, to_dict / to_pandas / to_polars exporters, missing_ok handling, calibrate_signal_array, Writer.add_reads, and per-read byte_count, bringing the escapepod bindings closer to the reference pod5 package.
• Python bindings for the signal layer: normalize, a KmerTable wrapper, and refine_signal_map are now exposed from escapepod-signal, so the resquiggle/normalization primitives are usable from Python.
• -t / --threads on filter, subset, and merge (matching demux) to cap the worker pool; these commands now default to a bounded pool of 8 threads instead of grabbing every CPU on a shared node.
• POD5 reads-table schema V5: the expected_open_pore_level and selected_read_level fields (both float32), introduced upstream in pod5 0.3.44, are now read, written, merged/filtered, surfaced in view/inspect (and as selectable output fields), and exposed on the Python ReadData / Writer.add_read API. Files escapepod writes are now stamped pod5_version 0.3.44 and verified readable by the reference ONT pod5 reader; existing V0–V4 files still read, with the new fields defaulting to 0.0.
• Defensive pre-mmap probe when opening a POD5 file: the header and footer are read through ordinary I/O before the file is memory-mapped, so a truncated file or an archive "stub" (unresident data on HSM/tape-backed filesystems) surfaces a recoverable error instead of an uncatchable SIGBUS on first page fault. Mirrors upstream pod5 0.3.37; set POD5_DISABLE_MMAP_OPEN=1 to skip.
• escpod demux <file> --model M -d OUT now runs a fused, streaming pipeline by default: each read's signal is decoded once, then detect + fingerprint + classify run in a single pass and reads are routed (block-level compressed copy) into per-barcode POD5s — no intermediate boundaries/fingerprints/classifications files. The granular detect/fingerprint/classify/split/train subcommands remain for advanced use. --classifications writes the per-read CSV only when asked.
• Experimental GPU primitives (behind --features gpu) for the demux signal chain — SVB16 decode, t-test fingerprint, LLR detect — kept as validated, reusable kernels. They are not used by escpod demux: measurement shows the prep stages run faster on a multi-core CPU than on the GPU.

Removed

• The GPU CNN adapter-detection path (demux detect --method cnn --gpu, plus the --cnn-weights flag and scripts/dump_adapter_cnn_weights.py). Its hand-written CUDA kernels were hardcoded to ADAPTed's BoundariesCNN topology (3× Conv1d + ConvTranspose1d, fixed K=7/C=64) and could not run any other architecture — including escapepod-models' replacement TCN. CNN detection (--method cnn) now runs only through the architecture-agnostic tract-onnx CPU path (adapter_cnn.rs), which accepts any [B,1,L] -> [B,2,L] ONNX graph. This is not a regression at typical scales: detect is dominated by POD5 read + signal prep, not CNN compute, so the CPU path is as fast or faster (the GPU flag's own help already said as much). Removing it also drops the CC-BY-NC .weights dumper. If a GPU CNN path is ever needed again, add an ONNX-generic backend (e.g. ORT CUDA EP) rather than a per-architecture kernel — which is exactly what the new cnn-gpu path (see Added) does.

Changed

• The CLI crate is renamed escapepod → escapepod-cli, matching its escapepod-{pod5,signal,demux} siblings and making its role explicit. The escpod binary name is unchanged, and installation is unchanged (cargo install --git …). Library consumers of the umbrella crate now import it as escapepod_cli (e.g. use escapepod_cli::signal) instead of escapepod. (The escapepod name on PyPI already belongs to the Python bindings, so this also removes the crate-name overlap.)
• CLI output split into logs vs. data. All status/progress/warning output now flows through tracing to stderr (timestamp LEVEL [target] message), while command data (TSV/CSV rows, inspect/summary reports, ID lists) stays on stdout, so it can be piped/redirected independently of logs. Default level is info; control it with -v/-vv/-q or RUST_LOG. Progress bars auto-hide under -q.
• Minimum supported Rust version raised to 1.95 (from 1.92).
• The resquiggle module is relicensed to MIT as an independent implementation, bringing it in line with the rest of the workspace.
• The Theil–Sen rescale subsample seed is now configurable (resquiggle --seed), making refinement runs bit-for-bit reproducible.
• The LLR detect --downscale default is now 10 (the WarpDemuX-native mode) for demux and demux detect, up from 1. This makes detect — the dominant prep stage — ~5× faster, with ~98% barcode agreement versus full-resolution (ds=1). Pass --downscale 1 to restore full-resolution detect.
• Dependency bumps (no behavior change): Arrow ecosystem arrow + parquet 58 → 59, tabled 0.20 → 0.21, the noodles-* BAM stack (bam 0.90, sam 0.85, bgzf 0.47, core 0.20, csi 0.56), ndarray 0.16 → 0.17, cudarc 0.12 → 0.19, and tract-onnx 0.21 → 0.23.

Performance

• Demux classify DTW is much faster, bit-identical. Per-call DTW allocation eliminated from the SVM classifier (~1.8×), then extended with lane-parallel SIMD DTW (8 signals/vector) covering windowed/penalty models as well — together a large end-to-end speedup on the DTW-bound path.
• GBM classify ~3.1× faster (compact tree arena + 8-read batched walk), bit-identical.
• Cold-read demux I/O fixed. Signal is now read in a single sequential memory-mapped stream with O(1) Arrow-batch seeks and dictionary pre-scan / readahead, instead of per-read random access across many threads (which degraded to demand-paging on cold files). Removes the large-POD5 startup stall and the cold-read throughput cliff.
• Zero-copy single-read signal fetch with adaptive prefix decode (#94): the reader's per-read signal path avoids materializing intermediate buffers, and fingerprint prep streams only the needed prefix.
• Faster Python read iteration: per-batch column resolution for the lazy for rd in reader: iterator and reads(), plus numpy-backed metadata columns for to_dict / to_pandas / to_polars.
• Parallelized bulk-file copies: filter, subset, and merge now copy reads across threads (bounded default pool of 8; -t to override), and the experimental index command adopts the same bounded default instead of inheriting rayon's all-CPU global pool.
• Single-pass demux split: reads are routed to per-barcode outputs in one pass, and the --unclassified flag now works correctly.
• Codebase-wide optimization/refactor sweep (#86), all bit-identical output:
  • Resquiggle adaptive banded DP ~31% faster — the per-base traceback no longer heap-allocates a Vec per base; the whole read shares one flat buffer.
  • O(1) POD5 read-batch access — read_batch(i) / read_ids_from_batch(i) now seek via the Arrow IPC footer instead of decoding every preceding batch. Iterating a many-read-batch file (e.g. the Python Reader read iterator) drops from O(B²) to O(B) batch decodes: ~10× faster random batch access and ~2.6× faster full-file iteration on a 1.65M-read / 166-batch file.
  • Signal median computations are O(n) instead of O(n log n) — the SVM kernel γ-heuristic, Theil–Sen rescale, and resquiggle dwell median now use select_nth_unstable instead of a full sort.
  • Smaller per-read allocations on the demux/classify and fingerprint hot paths (MAD-normalization scratch reuse, Platt coupling workspace sized once).
• Internal consolidation with no behavior change: six duplicated median impls unified into escapepod-signal::stats; the SVM RBF-kernel mapping and the CPU/GPU CNN batch packing/scatter each live in one shared helper.

Fixed

• Resolved a PyPI name collision: both the escapepod CLI crate and the escapepod-python bindings crate declared name = "escapepod". The PyPI escapepod distribution is the Python Reader bindings (escapepod-python); the escpod CLI now ships via cargo install escapepod and GitHub release binaries only, so its maturin pyproject.toml (a bindings = "bin" wheel) has been removed. This reverses the 0.5.1 note about pip install escapepod installing the CLI.

0.5.1 (2026-06-14)

Changed

• The CLI now ships from the escapepod crate (renamed from escapepod-cli), so cargo install escapepod installs the escpod binary. The same crate doubles as an umbrella library: with default-features = false plus pod5 / signal / demux, it re-exports each layer (e.g. escapepod::signal) without pulling in the CLI's dependency tree. demux stays opt-in until it stabilizes.
• The maturin binary wheel is published as escapepod (was escapepod-cli) so pip install escapepod matches cargo install.

Fixed

• Packaging readme pointed at a nonexistent path, which made cargo package fail; the workspace now points every publishable crate at the root README.md. escapepod-python is marked publish = false.
• demux fingerprint (test fixture): labeled-Parquet temp files lacked a .parquet suffix, so format detection read them as CSV and the parquet loaders failed with an "invalid UTF-8" error.

Build / Tooling

• Gated the train-only labeled-fingerprint loaders behind #[cfg(feature = "train")], removing dead-code warnings from --features demux builds.
• Bumped GitHub Actions to current majors (checkout v6, upload-artifact v7, download-artifact v8, setup-python v6, setup-pixi v0.9.6, action-gh-release v3, actions-netlify v4), clearing the Node 20 deprecation warnings.

0.5.0 (2026-04-27)

Added

• demux fingerprint: Parquet output when -o ends in .parquet, plus an --emit-dwell flag that adds per-segment dwell-time features.
• demux classify (CLI): fp_io module reads fingerprint inputs from both Parquet and CSV (gzipped CSV included); new flags --gpu-chunk-cells and --threads, with model auto-detection so --model accepts any supported format.
• escapepod-demux: AnyModel enum and load_any_model() for format-agnostic SVM/DTW model loading.
• escapepod-signal: SVM helper CUDA kernels exposed via function handles for downstream GPU pipelines.

Performance

• demux classify (GPU): on-GPU RBF + OvO decision pipeline (GpuSvmContext) keeps SVM evaluation on the device; producer/ consumer pipeline parallelizes the consumer side and bumps the default chunk to 4G cells with channel depth 2 for better GPU utilization on long runs. Per-chunk indicatif progress bar surfaces throughput.
• escapepod-demux: RBF kernel fast paths for power == 1.0 and power == 2.0 skip the generic powf call.
• escapepod-pod5: filter and merge hot paths reworked; remaining reader.reads() callers now batch-amortize the schema/footer parse, and a PoreType newtype removes per-read string churn.

Fixed

• train (multiclass OvO): dropped an unused SMO solve path that ran during training without contributing to the final model.

Build / Tooling

• Pixi dev env wires mold -run for fast local links (system gcc 11.5, no glibc-static needed); release artifacts in CI continue to build against musl.
• Docs: benchmark page leads with bulk operations (merge, filter, subset, repack); inspect and view demoted to a secondary section.

0.4.0 (2026-04-22)

Performance

• demux fingerprint: nested par_iter streams signals across files and reads so fingerprinting a 48-file run drops from ~32 min to ~9.8 s on the rna partition.
• demux classify / train-svm: reusable per-thread SvmWorkspace and a streaming (rayon fan-out + single writer) output path cut RSS by ~37% and remove a serialize-then-write stall.
• svb16: AVX2 decode path (16 samples/iter), preferred at runtime over SSSE3 when available.
• dtw: split the inner band loop so the trailing segment auto- vectorizes under AVX2; the x86-64-v3 baseline (AVX2 + FMA + BMI2 + POPCNT + F16C) is now pinned in .cargo/config.toml for portability across Broadwell/Cascade Lake/Ice Lake cluster nodes.
• segmentation::llr: allocation-free best_split and an opt-in early-stop variant.
• CLI: progress-bar updates throttled out of hot paths.

Changed

• Moved from CLI into libraries (additive for library consumers):
  • ReadBoundaries and fingerprint types/helpers now live in escapepod-demux.
  • normalize_signal(&[i16]) and the CLI's downscale_signal now live in escapepod-signal (the CLI's duplicate was removed).
• Docs: recommend srun -c 48 for throughput-sensitive demux runs on the rna partition (fills one socket without crossing NUMA).

0.3.1 (2026-04-21)

Added

• resquiggle::banded_dp_with_penalty_table — banded Viterbi DP variant that accepts a caller-supplied short-dwell penalty table and uses its length as the check horizon. Tie-break is strict (<), matching the Remora-compatible refinement pipeline. Complements the existing banded_dp which builds the asymmetric penalty internally.
• segmentation::mad_normalize_robust — median-MAD normalization with the 1.4826 Gaussian scale factor and graceful fallback (returns signal - median without dividing) when MAD is essentially zero. Complements mad_normalize, which panics on constant signals.

Performance

• Hot-path audit across fingerprint MAD, SVM prediction, view, and merge. Fingerprint MAD uses a single-pass median/MAD with an in-place partition; SVM prediction reuses per-thread scratch buffers and avoids redundant kernel evaluations on the OvO dual path; CLI view streams reads with reused formatting buffers; merge switches to mmap-backed readers where possible to cut per-file overhead.

Fixed

• escapepod-python cdylib now links cleanly under a plain cargo build on macOS. A build.rs emits the pyo3 extension-module link args (equivalent to pyo3_build_config::add_extension_module_link_args()), scoped to the cdylib, so the build no longer fails with undefined _Py* symbols when maturin is not driving the build. macOS is now in the CI matrix for check, test, and clippy to catch regressions.

Changed

• Workspace crates moved under crates/ (no public-API change).
• Docs reorganised with an "experimental tools" section; resquiggle and index CLI subcommands are gated behind their respective Cargo features.

0.3.0 (2026-04-20)

Breaking

• Barcode demultiplexing moved out of escapepod-signal into a new escapepod-demux crate. The escapepod_signal::demux module is gone; downstream code should depend on escapepod-demux directly and import from escapepod_demux::... (model loaders, classify_read, SVM predictor/trainer, Platt scaling, GPU batch classify, ADAPTed CNN adapter detection). The escpod demux CLI surface is unchanged, but escapepod-cli's demux Cargo feature now pulls in the new crate; the train, gpu, and cnn-detect features forward to it.

Added

• GPU-accelerated DTW for demux, opt-in via --features gpu on escapepod-signal and escapepod-cli. Wires up escpod demux classify --gpu (WarpDemuX model, CSV reference, and SVM model paths) and escpod demux train-svm --gpu. CUDA kernel is NVRTC-compiled at runtime, so no nvcc is required at build time — only the CUDA driver and libnvrtc at run time. On the lab cluster, pixi run -e gpu … provisions cuda-nvrtc via conda-forge and sets LD_LIBRARY_PATH. Anti-diagonal kernel with shared-memory-cached fingerprints; single- warp blocks with __syncwarp() and __launch_bounds__(32, 64). Measured ~7.7× speedup over CPU rayon on A30 at 1024×40×110 and 8192×40×110 workloads (criterion, band w=10).
• GpuDtwContext, dtw_distance_matrix_gpu, classify_reads_gpu, classify_with_svm_batch_gpu, compute_distance_matrix_gpu, train_svm_gpu public API on escapepod-signal (all gpu-gated).
• CPU classify_read now uses dtw_distance_bounded with the running second-best squared distance as an upper bound, skipping DTW work for training fingerprints that cannot change the top-2. Safe for both ratio and kernel threshold paths.

Fixed

• Behavior change for windowed DTW. The 2-row banded DP in dtw_distance / dtw_distance_bounded was leaving stale curr[j_start - 1] values from earlier rows, letting the recurrence read an out-of-band predecessor and occasionally find a shorter-than- valid path. The classical Sakoe-Chiba band treats those cells as unreachable; we now re-seed the boundary to INF at the top of each row and also short-circuit to INF when |n − m| > w (the endpoint itself is outside the band and the DP would otherwise propagate a stale in-band value through the trailing empty rows). Only affects callers that pass Some(window); unwindowed DTW is unchanged. In practice the difference is small but non-zero on real data — any downstream classify output produced with a band constraint may shift slightly, with GPU and CPU now agreeing bit-for-bit up to f32 summation order.

0.2.0 (2026-04-20)

Breaking

• Workspace split into two library crates. POD5 format I/O (reader, writer, VBZ compression, merge/filter/repack/subset, schema, footer, types, errors) now lives in the new escapepod-pod5 crate. The crate formerly called escapepod has been renamed to escapepod-signal and contains the signal-processing algorithms (DTW, resquiggle, segmentation) layered on top of escapepod-pod5. Downstream consumers depending on escapepod by name must rename to escapepod-signal; the pod5 surface is re-exported from escapepod-signal so most use escapepod::... paths translate to use escapepod_signal::... with no other changes.
• Barcode demultiplexing is now opt-in. The escapepod-signal::demux module and the escpod demux CLI subcommand require building with --features demux; the train feature now implies demux.

Added

• escapepod-pod5 crate for POD5 format I/O.
• demux Cargo feature on both escapepod-signal and escapepod-cli.

Changed

• README no longer advertises barcode demultiplexing as a shipped feature; docs/cli/demux.md carries an experimental admonition.
• CLI now declares demux and resquiggle commands as experimental in the commands index.

Removed

• Empty escapepod-vortex/ directory (content preserved on the escapepod-vortex branch).
• Stale PROGRESS.md and the top-level examples/test_ipc.rs scratch file; examples/dtw_example.rs moved under escapepod-signal/examples/.

0.1.3 (2026-04-20)

Added

• Tracing-based runtime verbosity (-v/-vv/-q, RUST_LOG)
• release-with-debug and profiling build profiles; phase timer
• Criterion microbenches covering audit hot paths (escapepod/benches/hot_paths.rs)

Changed

• SSSE3 SIMD encode/decode for SVB16 (~2× vs scalar on x86_64)
• Audit-driven hot-path optimizations across reader, DTW, demux, DP
• Dropped escapepod-vortex workspace member
• [profile.bench] pinned to inherit from release

Fixed

• Clippy lints: unnecessary_sort_by, needless_range_loop

0.1.0 (2026-03-19)

First stable release of escapepod-rs.

Added

• index: .p5i sidecar read index for fast UUID lookup (escpod index), with zstd-compressed entry blocks, sorted-vec binary search, and file size checksum validation
• filter: Sample count and end reason filters, stdin support for read IDs, fast reads_by_ids() path for UUID-only filtering
• subset: Accelerated subsetting via indexed batch lookup
• merge: Parallel I/O optimization

Fixed

• Include ZSTD content size in VBZ frames for Dorado/pod5 compatibility
• POD5 forward compatibility with Python pod5 library
• Correct pore count in summary table