chore: release v0.8.0

[nh13]

🤖 New release

• fgumi-bgzf: 0.7.0 -> 0.8.0
• fgumi-dna: 0.7.0 -> 0.8.0
• fgumi-tag: 0.7.0 -> 0.8.0
• fgumi-raw-bam: 0.7.0 -> 0.8.0
• fgumi-bam-io: 0.7.0 -> 0.8.0
• fgumi-cli-macros: 0.7.0 -> 0.8.0
• fgumi-metrics: 0.7.0 -> 0.8.0
• fgumi-sam: 0.7.0 -> 0.8.0
• fgumi-consensus: 0.7.0 -> 0.8.0
• fgumi-pipeline-core: 0.7.0 -> 0.8.0
• fgumi-sort: 0.7.0 -> 0.8.0
• fgumi-pipeline-io: 0.7.0 -> 0.8.0
• fgumi-simd-fastq: 0.7.0 -> 0.8.0
• fgumi-umi: 0.7.0 -> 0.8.0
• fgumi: 0.7.0 -> 0.8.0
• fgumi-fmt: 0.7.0 -> 0.8.0
• fgumi-cli-common: 0.7.0 -> 0.8.0

Changelog

fgumi-bgzf

[0.8.0] - 2026-10-06

Features

• Typed-step pipeline foundation (#870)
• Add --check-crc / --no-check-crc with the file-vs-stdin default (#933)

Refactor

• [breaking] Remove the legacy unified_pipeline engine (#947)

fgumi-dna

[0.8.0] - 2026-10-06

Performance

• Use fixed-seed ahash on hot per-item maps to remove RandomState global-counter contention (#865)

fgumi-tag

[0.8.0] - 2026-10-06

Bug Fixes

• [breaking] Give each methylation consensus caller one SEQ convention and make filters, simulate and docs follow it (#1010)
  • methylation::query_to_ref_positions takes the CIGAR as aligned (with its clips) and no longer takes the simplified CIGAR. * fix(consensus)!: emit observed bases from simplex methylation consensus Simplex EM-seq/TAPs consensus rewrote converted bases in the source reads (T->C / A->G at reference cytosines) before calling, so the emitted SEQ reported every site as unconverted. SEQ-based methylation callers and bisulfite-aware aligners read methylation from SEQ and saw every cytosine as methylated. The rewrite also hid real disagreements: all reads of a single-strand family copy one converted strand, so a C/T split is an error (or a UMI collision) and must lower quality like any other. Consensus is now always called on the observed bases, and simplex emits them as they are, with the evidence in cu/ct. MM/ML are no longer emitted for simplex: MM can only describe bases present in SEQ, and a converted cytosine is not one. A simplex consensus also cannot tell a converted cytosine from a C>T mutation, so restoring it would assert a base the molecule may not have. The shared MM builder now writes the SAM-spec '?' flag (unlisted bases are unknown), takes the conversion pattern from the read type (R1 C->T, R2 G->A) instead of the alignment strand, encodes each ML probability p as floor(256 p) capped at 255 (the SAM spec's [N/256, (N+1)/256) bins; it used floor(255 p), which put 1/2 and 3/4 in the bin below), and MN is available as a tag constant.
• Never auto-clip RG, MI and other non-per-base tags (#1019)

fgumi-raw-bam

[0.8.0] - 2026-10-06

Bug Fixes

• Replace deprecated wide swizzle_relaxed with shuffle (#992)
• [breaking] Give each methylation consensus caller one SEQ convention and make filters, simulate and docs follow it (#1010)
  • methylation::query_to_ref_positions takes the CIGAR as aligned (with its clips) and no longer takes the simplified CIGAR. * fix(consensus)!: emit observed bases from simplex methylation consensus Simplex EM-seq/TAPs consensus rewrote converted bases in the source reads (T->C / A->G at reference cytosines) before calling, so the emitted SEQ reported every site as unconverted. SEQ-based methylation callers and bisulfite-aware aligners read methylation from SEQ and saw every cytosine as methylated. The rewrite also hid real disagreements: all reads of a single-strand family copy one converted strand, so a C/T split is an error (or a UMI collision) and must lower quality like any other. Consensus is now always called on the observed bases, and simplex emits them as they are, with the evidence in cu/ct. MM/ML are no longer emitted for simplex: MM can only describe bases present in SEQ, and a converted cytosine is not one. A simplex consensus also cannot tell a converted cytosine from a C>T mutation, so restoring it would assert a base the molecule may not have. The shared MM builder now writes the SAM-spec '?' flag (unlisted bases are unknown), takes the conversion pattern from the read type (R1 C->T, R2 G->A) instead of the alignment strand, encodes each ML probability p as floor(256 p) capped at 255 (the SAM spec's [N/256, (N+1)/256) bins; it used floor(255 p), which put 1/2 and 3/4 in the bin below), and MN is available as a tag constant.

Features

• Typed-step pipeline foundation (#870)
• Declarative chain-builder layer + group --threads pilot (#872)

Performance

• Single-pass aux-tag rebuild (RawTagsEditor::rebuild_with) for zipper merge and NM/UQ/MD strip (#962)
• Fold per-record PG/AS/XS aux scans into single-pass walks (#964)
• Reduce serial-merge CPU and allocations at 1-2 threads (#971)

fgumi-bam-io

[0.8.0] - 2026-10-06

Bug Fixes

• Point empty stdin/pipe input at the upstream stage (#866)
• Return an error instead of panicking on >65,535 libraries (#956)

Features

• Typed-step pipeline foundation (#870)
• Inline BAI indexer on the arena sink (#883)
• Restore --read-streams on the arena chain via a concurrent scatter reader (#889)
• Add the bam-roundtrip subcommand (#936)

Performance

• Drop the discarded decode key and add skippable input CRC (#960)
• Fold UMI-position cache into the group-key aux scan (#976)

Refactor

• [breaking] Remove the legacy unified_pipeline engine (#947)

fgumi-cli-macros

[0.8.0] - 2026-10-06

fgumi-metrics

[0.8.0] - 2026-10-06

Documentation

• Stop calling the duplication ladder a saturation curve (#980)

Features

• Produce --metrics on the multi-threaded chain path (#915)
• Compute consensus QC metrics inline in fused runall and standalone consensus (#948)
• [breaking] Publish metric column manifest; headered filter --stats (#979)
  • fgumi filter --stats now writes a headered fgbio-Metric TSV (a total_reads/passed_reads/failed_reads/pass_rate header row plus one data row) instead of the previous headerless two-column key/value layout. Column names and semantics are unchanged; downstream parsers of the old vertical format must be updated. * test(metrics): round-trip conformance tests for all emitters * feat(metrics): publish metric column manifest with contract tests Adds crates/fgumi-metrics/metric_columns.json, the ordered column contract for every metric file fgumi emits, keyed . because group, simplex, duplex and dedup each write a differently shaped *.family_sizes.txt. tests/integration/test_metric_contract.rs parses the workspace with syn and fails in three cases: the manifest drifts from the live structs; a struct deriving serde's Serialize is neither listed nor explicitly allowlisted; or a serialized f64 field lacks the fgbio float encoding (Infinity/NaN rather than inf). downsample's --histogram-kept/--histogram-rejected were hand-formatted with writeln!. They now serialize a DownsampleHistogramMetric through the shared writer, so they are covered by the contract. The output is byte-identical. * fix(metrics): write headers for empty group and dedup histogram outputs group (--family-size-histogram, and the --metrics family- and position-group-size histograms) and dedup (--metrics, --family-size-histogram, --duplication-ladder) wrote through a bare DelimFile, which emits the header lazily. An empty histogram or ladder was therefore a 0-byte file that fgbio's Metric.read rejects ("No header found"). Route them through the shared write_metrics, which always writes the header. * docs(changelog): note filter --stats format break and metrics fixes * fix(metrics): honor .gz, special-file and symlink destinations in write_metrics write_metrics wrote rows to an extension-less temp file and renamed it into place. That had three problems. A .gz destination received plain text under a .gz name, which the metrics reader then failed to decompress. A non-regular destination such as /dev/stdout, a FIFO or a >(...) process substitution failed at rename time, after the whole run. A symlinked destination was replaced by a regular file. The temp file now carries the destination's extension so compression follows it. Non-regular destinations are written in place. Symlinks are resolved before the atomic rename. Empty outputs derive their header from a plain scratch file, so the gzip path keeps its header too.

Testing

• Fix inconsistent mate-strand flags in duplicate-group fixtures (#908)

fgumi-sam

[0.8.0] - 2026-10-06

Bug Fixes

• [breaking] Give each methylation consensus caller one SEQ convention and make filters, simulate and docs follow it (#1010)
  • methylation::query_to_ref_positions takes the CIGAR as aligned (with its clips) and no longer takes the simplified CIGAR. * fix(consensus)!: emit observed bases from simplex methylation consensus Simplex EM-seq/TAPs consensus rewrote converted bases in the source reads (T->C / A->G at reference cytosines) before calling, so the emitted SEQ reported every site as unconverted. SEQ-based methylation callers and bisulfite-aware aligners read methylation from SEQ and saw every cytosine as methylated. The rewrite also hid real disagreements: all reads of a single-strand family copy one converted strand, so a C/T split is an error (or a UMI collision) and must lower quality like any other. Consensus is now always called on the observed bases, and simplex emits them as they are, with the evidence in cu/ct. MM/ML are no longer emitted for simplex: MM can only describe bases present in SEQ, and a converted cytosine is not one. A simplex consensus also cannot tell a converted cytosine from a C>T mutation, so restoring it would assert a base the molecule may not have. The shared MM builder now writes the SAM-spec '?' flag (unlisted bases are unknown), takes the conversion pattern from the read type (R1 C->T, R2 G->A) instead of the alignment strand, encodes each ML probability p as floor(256 p) capped at 255 (the SAM spec's [N/256, (N+1)/256) bins; it used floor(255 p), which put 1/2 and 3/4 in the bin below), and MN is available as a tag constant.
• Keep MM/ML in step with hard-clipped and unmapped reads (#1012)
• Upgrade soft clips only at the mate-facing end when clipping overlaps (#1015)
• Never auto-clip RG, MI and other non-per-base tags (#1019)

Performance

• Single-pass aux-tag rebuild (RawTagsEditor::rebuild_with) for zipper merge and NM/UQ/MD strip (#962)

Testing

• Port missing fgbio SamRecordClipper and ClipBam tests (#1018)

fgumi-consensus

[0.8.0] - 2026-10-06

Bug Fixes

• [breaking] Reject duplex/codec flags the callers never honor (#997)
  • fgumi duplex --min-consensus-base-quality, fgumi codec --min-consensus-base-quality, fgumi codec --trim and their runall forms are removed. ConsensusCallingOptions loses trim and min_consensus_base_quality (now in QualityTrimOptions / MinConsensusBaseQualityOptions); DuplexOptions loses min_consensus_base_quality; CodecOptions and
• [breaking] Give each methylation consensus caller one SEQ convention and make filters, simulate and docs follow it (#1010)
  • methylation::query_to_ref_positions takes the CIGAR as aligned (with its clips) and no longer takes the simplified CIGAR. * fix(consensus)!: emit observed bases from simplex methylation consensus Simplex EM-seq/TAPs consensus rewrote converted bases in the source reads (T->C / A->G at reference cytosines) before calling, so the emitted SEQ reported every site as unconverted. SEQ-based methylation callers and bisulfite-aware aligners read methylation from SEQ and saw every cytosine as methylated. The rewrite also hid real disagreements: all reads of a single-strand family copy one converted strand, so a C/T split is an error (or a UMI collision) and must lower quality like any other. Consensus is now always called on the observed bases, and simplex emits them as they are, with the evidence in cu/ct. MM/ML are no longer emitted for simplex: MM can only describe bases present in SEQ, and a converted cytosine is not one. A simplex consensus also cannot tell a converted cytosine from a C>T mutation, so restoring it would assert a base the molecule may not have. The shared MM builder now writes the SAM-spec '?' flag (unlisted bases are unknown), takes the conversion pattern from the read type (R1 C->T, R2 G->A) instead of the alignment strand, encodes each ML probability p as floor(256 p) capped at 255 (the SAM spec's [N/256, (N+1)/256) bins; it used floor(255 p), which put 1/2 and 3/4 in the bin below), and MN is available as a tag constant.
• Keep MM/ML in step with hard-clipped and unmapped reads (#1012)

Features

• Typed-step pipeline foundation (#870)

Performance

• Use fixed-seed ahash on hot per-item maps to remove RandomState global-counter contention (#865)
• Single-pass scalar consensus-tag extraction (#969)
• Gated multi-base fast path for simplex consensus (#973)

Testing

• Speed up the slow-test tail without weakening invariants (#881)

fgumi-pipeline-core

[0.8.0] - 2026-10-06

Features

• Wire --pipeline-trace through the chain builder (#937)
• Thread-utilization and per-step bandwidth telemetry for the chain-builder engine (#951)
• Park idle pool workers to fix the thread-oversubscription cliff (#955)
• Add a refill-aware drain-first scheduler (#986)

Performance

• Don't fuse single-thread chains that declare Detached steps (#884)

Refactor

• [breaking] Remove the legacy unified_pipeline engine (#947)

Testing

• Read drained state once in pair-summer test steps (#981)

fgumi-sort

[0.8.0] - 2026-10-06

Bug Fixes

• [breaking] Enforce --max-temp-files on the arena spill path (#993)
  • fgumi-pipeline-io's SpillBlockEvent::Block gains a required key_kind field, and SpillWrite now emits SpillReady for its surviving runs at AllAnnounced (with slot_count set to their number) instead of as each run closes. RunMergerDyn::step takes a byte budget.
• Accept bwa's mid-pair -K split; emit BAM from the bwa-mem3 preset; keep mimalloc from purging (#988)
• Bound the mimalloc purge delay instead of never purging (#1017)

Documentation

• Remove in-tree design docs and gitignore docs/design/ (#934)

Features

• Typed-step pipeline foundation (#870)
• Declarative chain-builder layer + group --threads pilot (#872)
• Restore --read-streams on the arena chain via a concurrent scatter reader (#889)
• Give stage option structs a clap::Args surface for runall (PR A) (#907)
• Add the bam-roundtrip subcommand (#936)
• Thread-utilization and per-step bandwidth telemetry for the chain-builder engine (#951)
• Align in process with bwa-mem3 (--aligner::preset bwa-mem3-inproc) (#990)

Performance

• Use fixed-seed ahash on hot per-item maps to remove RandomState global-counter contention (#865)
• Extend runs in the arena spill path (t16 parity) (#886)
• Parallel gather for the in-memory sort fast path (#966)

Refactor

• [breaking] Remove the legacy unified_pipeline engine (#947)
• Stop configuring the chain sort through RawExternalSorter (#999)

fgumi-pipeline-io

[0.8.0] - 2026-10-06

Bug Fixes

• [breaking] Enforce --max-temp-files on the arena spill path (#993)
  • fgumi-pipeline-io's SpillBlockEvent::Block gains a required key_kind field, and SpillWrite now emits SpillReady for its surviving runs at AllAnnounced (with slot_count set to their number) instead of as each run closes. RunMergerDyn::step takes a byte budget.
• Leave no output files behind when a chain fails to build (#1006)

Features

• Typed-step pipeline foundation (#870)
• Inline BAI indexer on the arena sink (#883)
• Route the sort command onto the declarative chain builder (#885)
• Restore --read-streams on the arena chain via a concurrent scatter reader (#889)
• Thread --sort-stats diagnostics through the chain (#913)
• Add --check-crc / --no-check-crc with the file-vs-stdin default (#933)
• Expose --block-batch and --file-granularity as hidden flags (#943)
• Thread-utilization and per-step bandwidth telemetry for the chain-builder engine (#951)

Performance

• Don't fuse single-thread chains that declare Detached steps (#884)
• Extend runs in the arena spill path (t16 parity) (#886)
• Parallel gather for the in-memory sort fast path (#966)

Refactor

• [breaking] Remove the legacy unified_pipeline engine (#947)
• Stop configuring the chain sort through RawExternalSorter (#999)

Testing

• Speed up the slow-test tail without weakening invariants (#881)

fgumi-simd-fastq

[0.8.0] - 2026-10-06

fgumi-umi

[0.8.0] - 2026-10-06

Bug Fixes

• [breaking] Reverse the methylation counts with the Consensus tag set (#1011)
  • zipper --tags-to-reverse Consensus now also reverses cu, ct, au, at, bu and bt, and fgumi_umi::TagSets::CONSENSUS_REVERSE lists those six tags after fgbio's eight; a pipeline that reverses with both zipper Consensus and filter --reverse-per-base-tags now reverses those counts twice as well, so reverse in only one of the two.

Features

• Typed-step pipeline foundation (#870)
• Declarative chain-builder layer + group --threads pilot (#872)

Performance

• Use fixed-seed ahash on hot per-item maps to remove RandomState global-counter contention (#865)

fgumi

[0.8.0] - 2026-10-06

Bug Fixes

• Fall back to a name-only group key on an unreadable aux offset (#880)
• Thread --max-memory into the pipeline queue budget (#888)
• Resolve --max-temp-files once and bake it into the chain spec (#891)
• Filter secondary/supplementary reads by flag, not UNKNOWN_REF (#903)
• Group by molecule by default, add --per-strand opt-out (#904) (#906)
• Restore chain-path diagnostic parity (dedup index-threshold banner, copy_umi unwired-flags warning, pair_fastq docstring) (#914)
• Log the missing/wrong-length-UMI banner at error level on the chain path (#912)
• Error on --min-corrected with empty input on the legacy paths (parity with the chain/fgbio) (#919)
• Write --metrics only on success on the chain path (match clip/retag) (#921)
• Consult the UMI-position cache by assigning UMI groups before clearing duplicate flags (#918)
• Wire --methylation-mode/--ref into the fused filter stage (#944)
• Restore the === Sort Phase Timing === per-phase diagnostic (#945)
• Return an error instead of panicking on >65,535 libraries (#956)
• Honor --check-crc/--no-check-crc on the BGZF FASTQ split decode (#959)
• Replace deprecated wide swizzle_relaxed with shuffle (#992)
• Honor --correct::rejects when correct is fused mid-chain (#995)
• [breaking] Enforce --max-temp-files on the arena spill path (#993)
  • fgumi-pipeline-io's SpillBlockEvent::Block gains a required key_kind field, and SpillWrite now emits SpillReady for its surviving runs at AllAnnounced (with slot_count set to their number) instead of as each run closes. RunMergerDyn::step takes a byte budget.
• Honor --extract::no-check-crc on BGZF FASTQ input (#996)
• [breaking] Reject duplex/codec flags the callers never honor (#997)
  • fgumi duplex --min-consensus-base-quality, fgumi codec --min-consensus-base-quality, fgumi codec --trim and their runall forms are removed. ConsensusCallingOptions loses trim and min_consensus_base_quality (now in QualityTrimOptions / MinConsensusBaseQualityOptions); DuplexOptions loses min_consensus_base_quality; CodecOptions and
• [breaking] Reject the zipper flags the chain never reads (#998)
  • fgumi runall --zipper::buffer and
• [breaking] Enforce the shared consensus Phred bounds (#1001)
• Honor both async-reader flags on both FASTQ paths (#1000)
• [breaking] Reject --clipping-attribute, which extract never honors (#1004)
  • fgumi extract --clipping-attribute and fgumi runall --extract::clipping-attribute are removed. ExtractRunallOptions loses the public clipping_attribute field.
• Honor the --zipper:: merge rules on fused align chains (#1005)
• Leave no output files behind when a chain fails to build (#1006)
• Accept bwa's mid-pair -K split; emit BAM from the bwa-mem3 preset; keep mimalloc from purging (#988)
• [breaking] Store reverse-strand reads in reference orientation (#1008)
  • for a given seed, simulate mapped-reads, grouped-reads and consensus-reads now write reverse-strand SEQ and QUAL in reference orientation, soft-clip read-through padding when the insert is shorter than the read (changing CIGAR, MC and bin for those records), and consensus-reads writes reverse-strand per-base arrays reversed.
• [breaking] Give each methylation consensus caller one SEQ convention and make filters, simulate and docs follow it (#1010)
  • methylation::query_to_ref_positions takes the CIGAR as aligned (with its clips) and no longer takes the simplified CIGAR. * fix(consensus)!: emit observed bases from simplex methylation consensus Simplex EM-seq/TAPs consensus rewrote converted bases in the source reads (T->C / A->G at reference cytosines) before calling, so the emitted SEQ reported every site as unconverted. SEQ-based methylation callers and bisulfite-aware aligners read methylation from SEQ and saw every cytosine as methylated. The rewrite also hid real disagreements: all reads of a single-strand family copy one converted strand, so a C/T split is an error (or a UMI collision) and must lower quality like any other. Consensus is now always called on the observed bases, and simplex emits them as they are, with the evidence in cu/ct. MM/ML are no longer emitted for simplex: MM can only describe bases present in SEQ, and a converted cytosine is not one. A simplex consensus also cannot tell a converted cytosine from a C>T mutation, so restoring it would assert a base the molecule may not have. The shared MM builder now writes the SAM-spec '?' flag (unlisted bases are unknown), takes the conversion pattern from the read type (R1 C->T, R2 G->A) instead of the alignment strand, encodes each ML probability p as floor(256 p) capped at 255 (the SAM spec's [N/256, (N+1)/256) bins; it used floor(255 p), which put 1/2 and 3/4 in the bin below), and MN is available as a tag constant.
• [breaking] Reverse the methylation counts with the Consensus tag set (#1011)
  • zipper --tags-to-reverse Consensus now also reverses cu, ct, au, at, bu and bt, and fgumi_umi::TagSets::CONSENSUS_REVERSE lists those six tags after fgbio's eight; a pipeline that reverses with both zipper Consensus and filter --reverse-per-base-tags now reverses those counts twice as well, so reverse in only one of the two.
• Keep MM/ML in step with hard-clipped and unmapped reads (#1012)
• Label paired strands by read orientation in the parallel assigner (#1013)
• Bound the mimalloc purge delay instead of never purging (#1017)
• Never auto-clip RG, MI and other non-per-base tags (#1019)

Documentation

• Add NanoSeq (Duplex-Seq) pipeline guide (#910)
• Remove in-tree design docs and gitignore docs/design/ (#934)
• Document fgumi runall and add Reference guide pages (#938)
• Stop calling the duplication ladder a saturation curve (#980)
• Leave CHANGELOG.md to release-plz (#1003)

Features

• Add --threads and route through the unified pipeline (#871)
• Accept interleaved FASTQ input via --interleaved (#874)
• Typed-step pipeline foundation (#870)
• Warn when an explicit memory budget crowds the host (#867)
• Declarative chain-builder layer + group --threads pilot (#872)
• Wire dedup --threads onto the declarative chain builder (#876)
• Add command to copy the read-name UMI into the RX tag (#873)
• Inline BAI indexer on the arena sink (#883)
• Route the sort command onto the declarative chain builder (#885)
• Restore --read-streams on the arena chain via a concurrent scatter reader (#889)
• Route the correct command onto the declarative chain builder (#893)
• Route the filter command onto the declarative chain builder (#892)
• Route the simplex command onto the declarative chain builder (#894)
• Route the clip command onto the declarative chain builder (#897)
• Route the codec command onto the declarative chain builder (#895)
• Route the duplex command onto the declarative chain builder (#896)
• Route the retag command onto the declarative chain builder (#898)
• Route the copy-umi command onto the declarative chain builder (#899)
• Route the extract command onto the declarative chain builder (#900)
• Add opt-in --verify for strict template-coordinate order (#909)
• Thread --sort-stats diagnostics through the chain (#913)
• Give stage option structs a clap::Args surface for runall (PR A) (#907)
• Fused multi-stage fgumi runall command (#911)
• Produce --metrics on the multi-threaded chain path (#915)
• Add --check-crc / --no-check-crc with the file-vs-stdin default (#933)
• Add the aligner replay subcommand (#935)
• Add the bam-roundtrip subcommand (#936)
• Wire --pipeline-trace through the chain builder (#937)
• Add hidden --pool-scheduler override for A/B benchmarking (#942)
• Run BAM→FASTQ on the typed-step chain and add paired split output (#939)
• Expose --block-batch and --file-granularity as hidden flags (#943)
• Compute consensus QC metrics inline in fused runall and standalone consensus (#948)
• Thread-utilization and per-step bandwidth telemetry for the chain-builder engine (#951)
• Park idle pool workers to fix the thread-oversubscription cliff (#955)
• Add --threads to simplex-metrics and duplex-metrics (#968)
• [breaking] Publish metric column manifest; headered filter --stats (#979)
  • fgumi filter --stats now writes a headered fgbio-Metric TSV (a total_reads/passed_reads/failed_reads/pass_rate header row plus one data row) instead of the previous headerless two-column key/value layout. Column names and semantics are unchanged; downstream parsers of the old vertical format must be updated. * test(metrics): round-trip conformance tests for all emitters * feat(metrics): publish metric column manifest with contract tests Adds crates/fgumi-metrics/metric_columns.json, the ordered column contract for every metric file fgumi emits, keyed . because group, simplex, duplex and dedup each write a differently shaped *.family_sizes.txt. tests/integration/test_metric_contract.rs parses the workspace with syn and fails in three cases: the manifest drifts from the live structs; a struct deriving serde's Serialize is neither listed nor explicitly allowlisted; or a serialized f64 field lacks the fgbio float encoding (Infinity/NaN rather than inf). downsample's --histogram-kept/--histogram-rejected were hand-formatted with writeln!. They now serialize a DownsampleHistogramMetric through the shared writer, so they are covered by the contract. The output is byte-identical. * fix(metrics): write headers for empty group and dedup histogram outputs group (--family-size-histogram, and the --metrics family- and position-group-size histograms) and dedup (--metrics, --family-size-histogram, --duplication-ladder) wrote through a bare DelimFile, which emits the header lazily. An empty histogram or ladder was therefore a 0-byte file that fgbio's Metric.read rejects ("No header found"). Route them through the shared write_metrics, which always writes the header. * docs(changelog): note filter --stats format break and metrics fixes * fix(metrics): honor .gz, special-file and symlink destinations in write_metrics write_metrics wrote rows to an extension-less temp file and renamed it into place. That had three problems. A .gz destination received plain text under a .gz name, which the metrics reader then failed to decompress. A non-regular destination such as /dev/stdout, a FIFO or a >(...) process substitution failed at rename time, after the whole run. A symlinked destination was replaced by a regular file. The temp file now carries the destination's extension so compression follows it. Non-regular destinations are written in place. Symlinks are resolved before the atomic rename. Empty outputs derive their header from a plain scratch file, so the gzip path keeps its header too.
• [breaking] Add pair op to build a paired UMI from own/mate tags (#985)
  • RetagOp gains a Pair variant, and its public src() accessor is replaced by a crate-private sources().
• Add the aligner-bwa-mem3 feature, cohort math and the AlignEngine abstraction (#989)
• Align in process with bwa-mem3 (--aligner::preset bwa-mem3-inproc) (#990)
• Align EM-seq bisulfite-aware under --methylation-mode; bump bwa-mem3-rs to 0.4.2 (#1016)

Miscellaneous Tasks

• Mark the BGZF CRC verification policy as intentional (#932)
• Publish only when a push bumps the workspace version (#983)
• Bump codecov/codecov-action to v7.1.1 (#1020)

Performance

• Use fixed-seed ahash on hot per-item maps to remove RandomState global-counter contention (#865)
• Skip per-record group-key computation for key-agnostic groupers (#875)
• Build the merge tag bitsets once per batch, not per template (#917)
• Skip the discarded group key on the chain filter path (#916)
• Use DrainFirstScheduler for terminal grouping chains (#941)
• Single-pass aux-tag rebuild (RawTagsEditor::rebuild_with) for zipper merge and NM/UQ/MD strip (#962)
• Parallel queryname grouping for the correct/align/filter chains (#957)
• Resolve the template UMI from the primary read (#958)
• Fold per-record PG/AS/XS aux scans into single-pass walks (#964)
• Drop the discarded decode key and add skippable input CRC (#960)
• Decode-free BAM fast path for single-record filter (#961)
• Decode-free BAM fast path for copy-umi (#963)
• Parallel gather for the in-memory sort fast path (#966)
• Reduce per-record allocation (zero-copy segment views + reused Phred scratch) (#967)
• Decode-free BAM fast path for retag (#965)
• Single-pass scalar consensus-tag extraction (#969)
• Reduce serial-merge CPU and allocations at 1-2 threads (#971)
• Gated multi-base fast path for simplex consensus (#973)
• Scale the per-template merge past ~4.5 cores at 4+ threads (#975)
• Fold UMI-position cache into the group-key aux scan (#976)

Refactor

• Delete the CLI's throwaway sorter construction (#890)
• Retire the legacy single-threaded path; the chain is the only path (#920)
• Retire the legacy single-threaded path; the chain is the only path (#922)
• Retire the legacy single-threaded path; the chain is the only path (#923)
• Retire the legacy single-threaded path; the chain is the only path (#924)
• Retire the legacy single-threaded path; the chain is the only path (#925)
• [breaking] Collapse simplex/duplex/codec into a single consensus feature and retire the legacy single-threaded paths (#926)
• Retire the legacy single-threaded path; the chain is the only path (#927)
• Retire the legacy single-threaded paths; the chain is the only path (#928)
• Relocate reusable types out of unified_pipeline (R6a) (#929)
• [breaking] Remove the legacy unified_pipeline engine (#947)
• Split align+merge into a backend trait and a shared merge (#987)
• Stop configuring the chain sort through RawExternalSorter (#999)

Testing

• Speed up the slow-test tail without weakening invariants (#881)
• Pin TLEN + pair orientation against htsjdk across a broad matrix (#905)
• Fix inconsistent mate-strand flags in duplicate-group fixtures (#908)
• Consolidate the byte-identity file-compare idiom into a shared helper (#950)
• Close --all-metrics and group-metrics integration coverage gaps (#949)
• Shrink the BGZF no-check-crc fixture below the slow timeout (#1007)
• Port missing fgbio SamRecordClipper and ClipBam tests (#1018)

fgumi-fmt

[0.8.0] - 2026-10-06

fgumi-cli-common

[0.8.0] - 2026-10-06

---

This PR was generated with release-plz.

[coderabbitai]

Important

Review skipped

Ignore keyword(s) in the title.

⛔ Ignored keywords (2)

• chore(deps
• chore: release

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

• Configuration used: Repository: fulcrumgenomics/fgumi/.coderabbit.yaml
• Review profile: ASSERTIVE
• Plan: Essentials
• Run ID: 057a606e-4b8f-438f-bce6-db745c812fb2

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

Note

Reviews paused

Use the following commands to manage reviews:

• @coderabbitai resume to resume automatic reviews.
• @coderabbitai review to trigger a single review.

Use the checkboxes below for quick actions:

• ▶️ Resume reviews
• 🔍 Trigger review

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

_{Comment @coderabbitai help to get the list of available commands.}

[nh13]

@coderabbitai pause

[coderabbitai]

✅ Action performed

Reviews paused.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 96.45%. Comparing base (62b98c0) to head (8f479de).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #868      +/-   ##
==========================================
- Coverage   96.47%   96.45%   -0.03%     
==========================================
  Files         299      299              
  Lines      152214   152214              
==========================================
- Hits       146854   146818      -36     
- Misses       5360     5396      +36     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.