v2.8.0

A compatibility and correctness release. salmon now reads the versioned
RAD v2 format that current piscem writes, and the opt-in DAAREM EM
accelerator is reworked: it is now correct and fast with plain EM, and it is
no longer accepted with VBEM (the default optimizer). Nothing else about
quantification changes, and no index rebuild is needed: the index format
is unchanged from 2.7.0.

Results: unchanged by default

Against the released 2.7.0, on sample_data (paired-end, -l A), quant.sf
is byte-identical in every mode checked:

• default (VBEM)
• --useEM
• --sketch
• --emAccel squarem
• --useEM --emAccel daarem
• --numBootstraps 5

Indexes built by 2.7.0 and 2.8.0 are byte-identical at -p 1.

Reading RAD v2 (piscem ≥ 0.24)

piscem now writes RAD files in the versioned v2 format: a RAD_FILE
magic number, a spec version, a prelude extension block, and a semantic
role on every tag. salmon moves to libradicl 0.21, which detects the v2
prelude and reads bulk records by their declared roles. Quantifying piscem's
bulk RAD output (salmon quant on a RAD input) works with both the legacy
format and v2.

salmon's own RAD writer is unchanged and still writes the legacy format, so
its output is byte-identical to 2.7.0's. Emitting v2 from salmon is future
work.

--emAccel daarem: reworked, and plain EM only

Behavior change. --emAccel daarem now requires plain EM (--useEM or
--meta). With VBEM, the default optimizer, salmon exits with an error that
suggests --emAccel none or squarem.

Earlier releases could return inaccurate abundances with daarem:

• Negative intermediate abundances. The Anderson extrapolation
  overshot below zero and the negative iterate was kept. On a GENCODE run
  (646k transcripts) this produced a false "converged" far from the answer
  (Spearman 0.90 against a fully converged EM), with some reads lost to
  truncation.
• A convergence cycle under VBEM. The acceptance test judged steps by
  the residual ‖F(x) − x‖. VBEM's sparse prior drives near-identical
  isoforms to winner-take-all solutions, and the residual grows while the
  losing isoform drains. So the test refused VBEM's own path to the optimum:
  one run repeated the same state every 9,750 iterations until maxIter.

DAAREM now follows the objective-based variant of the reference method (the
CRAN daarem package's daarem_base_objfn):

• A step is accepted unless it lowers the log-likelihood by more than 0.01.
• Extrapolated abundances are floored so they stay positive.
• It stops only when both successive iterates and one plain EM step agree to
  within rel_diff_tol.
• Its vector work runs in parallel, and results are identical at any thread
  count.

With plain EM on three GENCODE v50 datasets (646k transcripts, 64 threads):

dataset	DAAREM, default tolerance	DAAREM, tolerance 1e-3	SQUAREM, tolerance 1e-4
SRR21186103	2.3 s	5.6 s (0.11 nats from optimum)	231 s
cerebellum (ERR13232583)	2.0 s	3.3 s (0.011)	153 s
testis (SRR11517405)	1.3 s	4.6 s (0.053)	313 s

Under VBEM, no DAAREM variant was an accelerator. It was slower than plain
VBEM at the default tolerance, and did not converge within 100k iterations at
1e-4 on two of the three datasets. VBEM's objective has many optima, and
the extrapolation kept crossing between them. With VBEM, use no acceleration
(the default) or --emAccel squarem, which pays off at tighter tolerances.

Library changes (salmon-infer)

• EmOptions::validate is new. It rejects DAAREM with VBEM, and the
  optimizer enforces it too.
• Target-less classes are skipped. An equivalence class with no targets
  is now ignored instead of causing an out-of-bounds panic. salmon never
  produces one, but external callers that build PackedEqClasses directly
  can (piscem-infer did, for fragments that failed its strand filter).
• SQUAREM docs corrected. They no longer promise the same fixed point
  as plain iteration under VBEM.

Dependencies

• libradicl 0.18 → 0.21. libradicl no longer depends on noodles.
• noodles unified on 0.116: noodles-bam 0.95, noodles-sam 0.90, and
  bramble-rs 0.1.9. The build now carries exactly one version of every
  noodles crate. This supersedes #1193 and #1194.
• Other updates: zstd 0.14, and patch updates to crossbeam, clap and jiff.

Also in this release

• Allocator code moved. The global allocator now lives in its own module
  (salmon-cli/src/global_alloc.rs), and the binary logs which allocator it
  was built with at debug level. docs/allocator-choice.md records what
  would be needed to justify changing the default (mimalloc). No behavior
  change. Thanks to @BenjaminDEMAILLE.
• Website. The documentation site builds again, and its dependencies were
  updated (Astro 7.3, Starlight 0.42).
• CI. GitHub Actions dependencies were updated.

Full details: docs/release-notes-2.8.0.md.

---

Install salmon-cli 2.8.0

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/COMBINE-lab/salmon/releases/download/v2.8.0/salmon-cli-installer.sh | sh

Download salmon-cli 2.8.0

File	Platform	Checksum
salmon-cli-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
salmon-cli-x86_64-apple-darwin.tar.xz	Intel macOS	checksum
salmon-cli-aarch64-unknown-linux-gnu.tar.xz	ARM64 Linux	checksum
salmon-cli-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum