benchmark: wire trim-galore-rs into the pipeline (#4)

trim-galore-rs was listed in performance.config.yaml and appeared
throughout RESULTS.md, but the supporting wiring had been lost between
checkouts: tools/trim-galore-rs/render.py was missing, and
config/tools.yaml had no entry. snakemake therefore expanded the trim
matrix to include trim-galore-rs jobs but only discovered the missing
render at job-dispatch time, after minutes of simulate / unrelated trim
work had already burned.

This commit restores the wiring and adds a guard so the same kind of
drift fails fast in the future:

- config/tools.yaml: add trim-galore-rs entry (display name, version
  2.2.0, version_cmd, pixi_env: trim-galore-rs).
- tools/trim-galore-rs/render.py: argv + output-name moves for the
  Rust trim_galore. Drops the Perl render's 4-core cap since the Rust
  binary uses an in-process thread pool that scales past 4.
- pixi.toml: new trim-galore-rs feature pinning trim-galore = "2.2.0.*"
  in its own environment + solve group, mirroring the fastp-nfcore
  pattern. Scoped to platforms with a bioconda 2.x build
  (linux-64 / linux-aarch64 / osx-arm64); osx-64 stays on Perl 0.6.x.
- install.sh: pixi install --environment trim-galore-rs; skipped on
  osx-64 where no 2.x binary exists.
- config/smoke.config.yaml: add trim-galore-rs to the smoke tool list
  so render.py / pixi env drift is caught in the ~15 min smoke run.
- Snakefile: validate at config-load time that every tool in
  config['tools'] has a tools/<render_dir>/render.py, raising
  WorkflowError immediately instead of failing partway through a sweep.

Smoke run on osx-arm64 (M1 Max, 4 threads, 1M PE pairs/sample) shows
trim-galore-rs landing in the expected position relative to chelae,
cutadapt, and fastp on both insert-size shapes.

Author: tfenne

benchmark-pipeline/Snakefile

@@ -105,6 +105,20 @@ TOOL_NAMES = config["tools"]
 THREAD_COUNTS = config["thread_counts"]
 REPLICATES = list(range(1, int(config["replicates"]) + 1))
 
+# Fail fast if any tool in the config lacks a render.py. Without this
+# guard the missing render only surfaces when `trim_one` actually runs
+# for that tool, which can be many minutes into a sweep (see issue #4).
+_missing_renders = [
+    t for t in TOOL_NAMES
+    if not (WORKDIR / "tools" / TOOLS.get(t, {}).get("render_dir", t) / "render.py").exists()
+]
+if _missing_renders:
+    raise WorkflowError(
+        "Tools listed in config['tools'] have no render.py: "
+        f"{_missing_renders}. Add tools/<name>/render.py and a "
+        "matching entry to config/tools.yaml."
+    )
+
 
 def all_trim_outputs():
     out = []

benchmark-pipeline/config/smoke.config.yaml

@@ -51,6 +51,7 @@ tools:
   - fastp-nfcore
   - cutadapt
   - trim-galore
+  - trim-galore-rs
   - trimmomatic
   - bbduk
   - adapterremoval

benchmark-pipeline/config/tools.yaml

@@ -48,6 +48,18 @@ trim-galore:
   version_cmd: "trim_galore --version | head -n 5"
   source: bioconda-latest
 
+# Trim Galore 2.x (Rust rewrite, "Oxidized Edition"). Shares the
+# `trim-galore` bioconda package name with the Perl version pinned at
+# 0.6.11 in the default env, so it gets its own pixi env. The render
+# directory differs from the Perl render — the Rust binary scales past
+# 4 threads so we don't apply the Perl render's cores cap.
+trim-galore-rs:
+  display_name: "TrimGalore 2.x (Rust)"
+  version: "2.2.0"
+  version_cmd: "trim_galore --version | head -n 5"
+  pixi_env: trim-galore-rs
+  source: bioconda-latest
+
 trimmomatic:
   display_name: Trimmomatic
   version: "0.40"

benchmark-pipeline/install.sh

@@ -88,11 +88,19 @@ else
 fi
 
 # ---- pixi environments ----------------------------------------------------
-log "Materializing pixi environments (default, run, plot, fastp-nfcore)"
+log "Materializing pixi environments (default, run, plot, fastp-nfcore, trim-galore-rs)"
 pixi install                              # default
 pixi install --environment run
 pixi install --environment plot
 pixi install --environment fastp-nfcore
+# trim-galore-rs is only built for linux-64 / linux-aarch64 / osx-arm64 on
+# bioconda; on osx-64 the feature's `platforms` restriction makes the env
+# unsolvable. Skip the install there with a warning rather than aborting.
+if [[ "$(uname -s)-$(uname -m)" == "Darwin-x86_64" ]]; then
+  log "Skipping pixi env 'trim-galore-rs' on osx-64 (no bioconda 2.x build)"
+else
+  pixi install --environment trim-galore-rs
+fi
 
 # ---- rust toolchain -------------------------------------------------------
 RUST_DIR="$PIPELINE_DIR/.rust"

benchmark-pipeline/pixi.lock

@@ -51,8 +51,21 @@ r-scales = "*"
 [feature.fastp-nfcore.dependencies]
 fastp = "1.1.0.*"
 
+# Trim Galore 2.x (Oxidized Edition): Rust rewrite of the Perl trim_galore,
+# shares the `trim-galore` bioconda package name with the Perl 0.6.x line.
+# Own environment so it doesn't collide with the trim-galore = "0.6.11.*"
+# pin in the default trimmers env. osx-64 has no 2.x build on bioconda
+# (latest there is 0.6.2), so the feature is scoped to the platforms that
+# actually have a 2.x package.
+[feature.trim-galore-rs]
+platforms = ["linux-64", "linux-aarch64", "osx-arm64"]
+
+[feature.trim-galore-rs.dependencies]
+trim-galore = "2.2.0.*"
+
 [environments]
-default      = { features = ["trimmers", "plotting"], solve-group = "main" }
-run          = { features = ["trimmers"],              solve-group = "main" }
-plot         = { features = ["plotting"],              solve-group = "main" }
-fastp-nfcore = { features = ["fastp-nfcore"], no-default-feature = true, solve-group = "fastp-nfcore-solve" }
+default        = { features = ["trimmers", "plotting"], solve-group = "main" }
+run            = { features = ["trimmers"],              solve-group = "main" }
+plot           = { features = ["plotting"],              solve-group = "main" }
+fastp-nfcore   = { features = ["fastp-nfcore"],   no-default-feature = true, solve-group = "fastp-nfcore-solve" }
+trim-galore-rs = { features = ["trim-galore-rs"], no-default-feature = true, solve-group = "trim-galore-rs-solve" }

benchmark-pipeline/pixi.toml

@@ -0,0 +1,54 @@
+"""Trim Galore Oxidized Edition (Rust rewrite). CLI is a near-superset of
+the Perl trim_galore: same `--paired`, `--adapter`, `--adapter2`,
+`--stringency`, `--quality`, `--length`, `--output_dir`, `--cores`,
+`--gzip`. Output filenames also follow the Perl convention
+(`<basename>_val_1.fq.gz` for paired, `<basename>_trimmed.fq.gz` for SE),
+so the moves table reuses the same logic as the Perl render.
+
+Difference from the Perl render: no `--cores` cap at 4. The Rust binary
+uses an in-process thread pool rather than the Perl version's multi-
+process spawn, so it scales cleanly past 4 threads."""
+
+from pathlib import Path
+
+
+def render(ctx: dict) -> dict:
+    cfg = ctx["trim_cfg"]
+    workdir = Path(ctx["workdir"])
+
+    stringency = str(cfg.get("min_adapter_overlap", 5))
+    argv = ["trim_galore",
+            "--cores", str(ctx["threads"]),
+            "--adapter", ctx["adapter_r1"],
+            "--stringency", stringency,
+            "--gzip",
+            "--output_dir", str(workdir)]
+
+    if ctx["paired"]:
+        argv += ["--paired", "--adapter2", ctx["adapter_r2"]]
+
+    argv += ["--quality", str(cfg.get("quality_threshold", 0)) if cfg.get("quality_trim") else "0"]
+    argv += ["--length", str(cfg["min_length"]) if cfg.get("min_length", 0) > 0 else "1"]
+
+    argv.append(ctx["input_r1"])
+    if ctx["paired"]:
+        argv.append(ctx["input_r2"])
+
+    def basename_no_fqgz(path: str) -> str:
+        n = Path(path).name
+        for suffix in (".fastq.gz", ".fq.gz", ".fastq", ".fq"):
+            if n.endswith(suffix):
+                return n[: -len(suffix)]
+        return Path(path).stem
+
+    moves: dict[str, str] = {}
+    if ctx["paired"]:
+        b1 = basename_no_fqgz(ctx["input_r1"])
+        b2 = basename_no_fqgz(ctx["input_r2"])
+        moves[str(workdir / f"{b1}_val_1.fq.gz")] = ctx["output_r1"]
+        moves[str(workdir / f"{b2}_val_2.fq.gz")] = ctx["output_r2"]
+    else:
+        b1 = basename_no_fqgz(ctx["input_r1"])
+        moves[str(workdir / f"{b1}_trimmed.fq.gz")] = ctx["output_r1"]
+
+    return {"argv": argv, "moves": moves}