Fixes add yamls

scripts/run_benchmark/param_sweep/baysor_params.yaml

@@ -0,0 +1,64 @@
+# Parameter sweep for the baysor transcript-assignment method.
+# Committed source of truth, read at runtime from GitHub via a raw URL by
+# run_test_baysor_nebius.sh (the Nebius compute env pulls the repo but cannot see the
+# launch host's local files).
+#
+# Consumed by the run_benchmark workflow via the `method_parameters_yaml` setting
+# (src/workflows/run_benchmark/main.nf). For every method the workflow builds:
+#   * one "default" variant using the `default:` args below, and
+#   * one extra variant per value in each `sweep:` list, with that ONE arg overridden.
+# The benchmark varies a SINGLE parameter at a time (a "star" around the default, not
+# a full grid), so total baysor variants = 1 default + sum(sweep list lengths) = 11.
+#
+# See src/methods_transcript_assignment/baysor/NOTES.md ("Optimization / tuning") for the
+# tiers and rationale.
+#
+# WHAT BAYSOR ACTUALLY VARIES HERE: baysor runs Baysor v0.7.1 through sopa
+# (sopa.segmentation.baysor). It re-segments transcripts starting from the upstream
+# segmentation, fed in as a per-transcript prior. Coordinates are in ~1 um/px global space, so
+# `scale` (Baysor's typical cell radius) is effectively in um. `scale = -1.0` is the sentinel
+# for "estimate the radius from min_molecules_per_cell".
+#
+# NON-DEFAULT AUDIT (config default -> Baysor/sopa true default):
+#   * prior_segmentation_confidence: config 0.8 DEVIATES from Baysor default 0.2 (Baysor only
+#     recommends >0.7 when the prior is highly reliable) -> forced onto a sweep axis, walking
+#     back toward the tool default (0.2, and sopa's example 0.5).
+#   * min_molecules_per_cell: config 50 DEVIATES from sopa's example 10 (Baysor has no hard
+#     default) -> forced onto a sweep axis toward the smaller value.
+#   * scale: config -1.0 (auto-estimate) is a deliberate mode, not a tool default; swept to
+#     explicit um radii straddling sopa's example 6.25.
+#   * scale_std ("25%") and n_clusters (4) MATCH the Baysor defaults; probed lightly as Tier-2
+#     companion axes (default value omitted from each list, as required).
+#   * force_2d (true) matches sopa's default and the data is 2D -> left fixed (flipping to 3D is
+#     meaningless here). transcripts_key / coordinate_system are structural, not swept.
+#     Perf/resource knobs (patch_width/patch_overlap, JULIA_NUM_THREADS) are hardcoded, not swept.
+#
+# SUBMITTABLE-NOW: the Nebius launch runs `--revision build/main`, and every swept arg below is
+# ALREADY exposed in src/methods_transcript_assignment/baysor/config.vsh.yaml today, so the
+# build/main container accepts them with NO rebuild.
+parameters:
+  baysor:
+    # Baseline == the component's shipped defaults.
+    default:
+      force_2d: true
+      min_molecules_per_cell: 50
+      scale: -1.0
+      scale_std: "25%"
+      n_clusters: 4
+      prior_segmentation_confidence: 0.8
+    sweep:
+      # ---- Tier 1: cell-size / recall levers (sharpest) ----
+      # scale: target cell radius in um (coords ~1 um/px). Default -1.0 = auto-estimate.
+      # Explicit radii straddling sopa's example 6.25 um; grounded in typical cell radii ~4-10 um.
+      scale: [5.0, 7.5, 10.0]
+      # min_molecules_per_cell: min molecules to call a cell; also drives the auto scale estimate.
+      # DEVIATES (config 50 vs sopa 10) -> walk toward the smaller, more permissive value.
+      min_molecules_per_cell: [10, 25]
+      # prior_segmentation_confidence: trust the prior vs re-segment (in [0,1]).
+      # DEVIATES (config 0.8 vs Baysor 0.2) -> walk back toward the tool default and sopa's 0.5.
+      prior_segmentation_confidence: [0.2, 0.5]
+      # ---- Tier 2: companion knobs at the Baysor default ----
+      # scale_std: prior std of cell radius, relative to scale. Default "25%" -> wider tolerance.
+      scale_std: ["50%"]
+      # n_clusters: number of molecule clusters (~ major cell types). Default 4; Baysor rec. 3-15.
+      n_clusters: [6, 8]

scripts/run_benchmark/param_sweep/clustermap_params.yaml

@@ -0,0 +1,75 @@
+# Parameter sweep for the clustermap (ClusterMap, He et al. 2021) transcript-assignment method.
+# Committed source of truth, read at runtime from GitHub via a raw URL by
+# run_test_clustermap_nebius.sh (the Nebius compute env pulls the repo but cannot see the
+# launch host's local files).
+#
+# Consumed by the run_benchmark workflow via the `method_parameters_yaml` setting
+# (src/workflows/run_benchmark/main.nf). For every method the workflow builds:
+#   * one "default" variant using the `default:` args below, and
+#   * one extra variant per value in each `sweep:` list, with that ONE arg overridden.
+# The benchmark varies a SINGLE parameter at a time (a "star" around the default, not a full
+# grid), so total clustermap variants = 1 default + sum(sweep list lengths) = 14.
+#
+# See src/methods_transcript_assignment/clustermap/NOTES.md ("Optimization / tuning") for the
+# tiers, the ClusterMap-source audit, and the inert-arg findings.
+#
+# WHAT CLUSTERMAP ACTUALLY VARIES HERE: ClusterMap is segmentation-free — it density-peak-
+# clusters the RNA spots (+ injected DAPI points) into cells. The swept knobs are the levers
+# that reach the tool AND change the output, confirmed by reading script.py and the upstream
+# ClusterMap source:
+#   xy_radius (spot-neighbourhood radius, px), cell_num_threshold (DPC cell-count threshold),
+#   dapi_grid_interval (DAPI-point density), min_spot_per_cell (small-cell filter),
+#   pct_filter (low-density spot removal), gauss_blur (DAPI blur on/off).
+#
+# NON-DEFAULT AUDIT vs ClusterMap true defaults: three config defaults DEVIATE from the tool
+# and are therefore forced onto sweep axes (each list walks that arg back TOWARD its tool
+# default, since the default variant already covers the shipped value):
+#   gauss_blur         True -> tool False  (sweep [false])
+#   pct_filter         0.0  -> tool 0.1    (sweep toward 0.1)
+#   cell_num_threshold 0.1  -> tool 0.01   (sweep toward 0.01)
+# A fourth deviation, `contamination` (0 -> tool 0.1), is DEAD CODE while LOF=False (the shipped
+# default) and sklearn's LocalOutlierFactor rejects contamination=0, so LOF/contamination form a
+# coupled, unsafe pair in the star model and are NOT swept (see NOTES.md). `window_size` (700) is
+# a memory/perf tiling knob, not a ClusterMap parameter, so it is left fixed.
+#
+# INERT config args EXCLUDED from the sweep (they never reach ClusterMap): `use_dapi` (never read
+# by script.py), `add_dapi` and `use_genedis` (both overwritten by par["xy_radius"] via a script
+# bug). Sweeping any of them would just produce duplicate variants. Documented in NOTES.md.
+#
+# SUBMITTABLE-NOW: the Nebius launch runs `--revision build/main`, so the sweep touches ONLY args
+# already exposed + correctly forwarded in the build/main container -- all six below exist in
+# src/methods_transcript_assignment/clustermap/config.vsh.yaml today, so no rebuild is needed.
+parameters:
+  clustermap:
+    # Baseline == the component's shipped defaults (the "default variant").
+    default:
+      window_size: 700
+      xy_radius: 40
+      z_radius: 0
+      fast_preprocess: false
+      gauss_blur: true
+      sigma: 1.0
+      pct_filter: 0.0
+      LOF: false
+      contamination: 0
+      min_spot_per_cell: 5
+      dapi_grid_interval: 5
+      cell_num_threshold: 0.1
+    sweep:
+      # ---- Tier 1: sharpest quality levers ----
+      # xy_radius: spot-neighbourhood radius in PIXELS (~1 um/px grid). Default 40 is large vs
+      # typical cell radii; straddle tighter and looser neighbourhoods.
+      xy_radius: [20, 30, 60]
+      # cell_num_threshold: DPC cell-count threshold, "larger -> more cells". DEVIATION (0.1 vs
+      # tool 0.01); 0.01 = tool default (fewer/larger cells), 0.2 = more/smaller cells.
+      cell_num_threshold: [0.01, 0.05, 0.2]
+      # dapi_grid_interval: density of injected DAPI points; smaller = denser/slower.
+      dapi_grid_interval: [3, 10]
+      # min_spot_per_cell: min transcripts to keep a cell; higher deletes small clusters.
+      min_spot_per_cell: [10, 20]
+      # ---- Tier 2: quality knobs / must-sweep deviations ----
+      # pct_filter: low-density spot removal. DEVIATION (0.0 removes nothing vs tool 0.1); walk up.
+      pct_filter: [0.05, 0.1]
+      # gauss_blur: DAPI blur before binarising. DEVIATION (True vs tool False); boolean -> the
+      # single non-default value is the tool default False.
+      gauss_blur: [false]

scripts/run_benchmark/param_sweep/comseg_params.yaml

@@ -0,0 +1,74 @@
+# Parameter sweep for the comseg (ComSeg, Python/sopa) transcript-assignment method.
+# Committed source of truth, read at runtime from GitHub via a raw URL by
+# run_test_comseg_nebius.sh (the Nebius compute env pulls the repo but cannot see the
+# launch host's local files).
+#
+# Consumed by the run_benchmark workflow via the `method_parameters_yaml` setting
+# (src/workflows/run_benchmark/main.nf). For every method the workflow builds:
+#   * one "default" variant using the `default:` args below, and
+#   * one extra variant per value in each `sweep:` list, with that ONE arg overridden.
+# The benchmark varies a SINGLE parameter at a time (a "star" around the default, not
+# a full grid), so total comseg variants = 1 default + sum(sweep list lengths) = 13.
+#
+# See src/methods_transcript_assignment/comseg/NOTES.md ("Optimization / tuning") for the
+# tiers and rationale.
+#
+# WHAT COMSEG ACTUALLY VARIES HERE: comseg runs sopa's ComSeg wrapper, which hands a config
+# dict to the ComSeg point-cloud clustering algorithm (Defard et al., Comms Biol 2024). The
+# component exposes exactly six SCIENCE knobs of that dict (mean_cell_diameter, max_cell_radius,
+# alpha, min_rna_per_cell, norm_vector, allow_disconnected_polygon); the other config knobs
+# (patch_width/overlap/transcript_patch_width, n_workers, worker_memory_limit) are memory/perf
+# and resource knobs, and gene_column/transcripts_key/coordinate_system are structural, so none
+# of those is swept.
+#
+# NON-DEFAULT AUDIT (vs sopa's own _get_default_config, the tool baseline): TWO component
+# defaults deviate from the tool default and neither is a perf/resource knob, so each is FORCED
+# onto a sweep axis --
+#   * allow_disconnected_polygon: component TRUE vs sopa False -> sweep the flip (false).
+#   * min_rna_per_cell:           component 5    vs sopa 20    -> sweep walks 5 back toward 20.
+# The other four (mean_cell_diameter=15, max_cell_radius=25, alpha=0.5, norm_vector=false) sit at
+# or near sopa's defaults but are the algorithm's core science levers, so they are swept too.
+#
+# SUBMITTABLE-NOW: the Nebius launch runs `--revision build/main`, so the sweep touches ONLY
+# args already exposed in the build/main container -- all six knobs below already exist in
+# src/methods_transcript_assignment/comseg/config.vsh.yaml today, so NO rebuild is needed.
+# ComSeg's deeper clustering knobs (co-expression graph n_neighbors, k_nearest_neighbors,
+# leiden resolution, min_nb_rna_patch) are NOT plumbed through the component config dict at all
+# (ComSeg/sopa use their own defaults); exposing them would need a config arg + container
+# rebuild, so they are documented in NOTES.md as future work and NOT here.
+parameters:
+  comseg:
+    # Baseline == the component's shipped defaults (the six exposed ComSeg science knobs).
+    default:
+      mean_cell_diameter: 15.0
+      max_cell_radius: 25.0
+      alpha: 0.5
+      min_rna_per_cell: 5
+      norm_vector: false
+      allow_disconnected_polygon: true
+    sweep:
+      # ---- Tier 1: cell-size geometry (µm), the sharpest levers on what gets grouped ----
+      # mean_cell_diameter: expected mean cell diameter in µm; sets the graph scale ComSeg
+      # builds neighborhoods on. Shipped 15µm (a typical mammalian cell). Straddle smaller /
+      # larger tissue cells. (sopa's true default DERIVES this from the prior-segmentation
+      # areas; the component instead pins 15µm.)
+      mean_cell_diameter: [10.0, 20.0]
+      # max_cell_radius: max centroid->RNA distance (µm) for association; caps how far a cell
+      # can reach out to claim transcripts. Shipped 25µm; sopa's rule-of-thumb is
+      # mean_cell_diameter*1.75 (~26µm at 15µm), the ComSeg docs example uses 50µm. Walk from
+      # tighter (20) up toward the docs value (50).
+      max_cell_radius: [20.0, 37.5, 50.0]
+      # alpha: alphashape parameter (0..1) for the cell polygon; 1 == convex hull, lower ==
+      # tighter/more concave boundary hugging the points. Shipped 0.5. Sweep tighter, looser,
+      # and full convex hull.
+      alpha: [0.25, 0.75, 1.0]
+      # min_rna_per_cell: minimum transcripts for a cell to be kept. Shipped 5 (relaxed);
+      # sopa's default is 20. NON-DEFAULT AUDIT -> walk 5 back toward the sopa default 20
+      # (stricter = fewer, more confident cells).
+      min_rna_per_cell: [10, 20]
+      # norm_vector (bool): normalize the per-cell expression vectors before the co-expression
+      # graph / clustering. Shipped false (== sopa default). Boolean -> one flip.
+      norm_vector: [true]
+      # allow_disconnected_polygon (bool): allow a cell's boundary to be a multi-part polygon.
+      # Shipped true; sopa's default is false. NON-DEFAULT AUDIT -> sweep the flip (false).
+      allow_disconnected_polygon: [false]

scripts/run_benchmark/param_sweep/fastreseg_params.yaml

@@ -0,0 +1,65 @@
+# Parameter sweep for the fastreseg (FastReseg, R) transcript-assignment method.
+# Committed source of truth, read at runtime from GitHub via a raw URL by
+# run_test_fastreseg_nebius.sh (the Nebius compute env pulls the repo but cannot see the
+# launch host's local files).
+#
+# Consumed by the run_benchmark workflow via the `method_parameters_yaml` setting
+# (src/workflows/run_benchmark/main.nf). For every method the workflow builds:
+#   * one "default" variant using the `default:` args below, and
+#   * one extra variant per value in each `sweep:` list, with that ONE arg overridden.
+# The benchmark varies a SINGLE parameter at a time (a "star" around the default, not
+# a full grid), so total fastreseg variants = 1 default + sum(sweep list lengths) = 11.
+#
+# See src/methods_transcript_assignment/fastreseg/NOTES.md ("Optimization / tuning") for the
+# tiers and rationale.
+#
+# !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
+# NEEDS REBUILD (unlike the other six sibling sweeps, this one is NOT submittable as-is):
+# these four knobs did NOT exist as Viash args until this change — they were HARDCODED inside
+# script.R's fastReseg_full_pipeline(...) call. Exposing them adds four `--...` args to
+# config.vsh.yaml + wires them through orchestrator.sh -> script.R. The `build/main` container
+# run by run_test_fastreseg_nebius.sh's `--revision build/main` will REJECT these args until
+# the component is rebuilt. Before launching you MUST:
+#   1. commit + push this file AND the config/orchestrator/script.R edits,
+#   2. `viash ns build` (regenerate target/) and rebuild the fastreseg container on ghcr
+#      with the build_main tag (see the check-component skill), then
+#   3. launch. Otherwise every fastreseg variant fails with an "unknown option" error.
+# !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
+#
+# NON-DEFAULT AUDIT: every hardcoded value in script.R exactly matched FastReseg's TRUE package
+# default for fastReseg_full_pipeline() (pixel_size 0.18, zstep_size 0.8,
+# molecular_distance_cutoff 2.7, flagCell_lrtest_cutoff 5, svmClass_score_cutoff -2,
+# groupTranscripts_method "dbscan", spatialMergeCheck_method "leidenCut", cutoff_spatialMerge
+# 0.5, all NULL cutoffs). So NONE was a pre-tuned deviation demanding a sweep axis; the config
+# defaults below are set to those same package defaults and the sweep walks each knob AWAY from
+# its default in both directions.
+#
+# NOT SWEPT (documented in NOTES.md): pixel_size / zstep_size are Tier-0 dataset-physical
+# geometry constants (still hardcoded, not exposed); groupTranscripts_method /
+# spatialMergeCheck_method are Tier-2/3 categorical toggles (hardcoded). The four swept below
+# are the Tier-1 error-detection / correction levers.
+parameters:
+  fastreseg:
+    # Baseline == FastReseg's own package defaults (previously hardcoded in script.R).
+    default:
+      molecular_distance_cutoff: 2.7
+      flagCell_lrtest_cutoff: 5
+      svmClass_score_cutoff: -2
+      cutoff_spatialMerge: 0.5
+    sweep:
+      # ---- Tier 1: molecular neighborhood distance (SHARPEST lever) ----
+      # Max molecule-to-molecule distance (um) for grouping transcripts into candidate cells.
+      # Default 2.7; smaller => tighter/more fragmented groups, larger => looser groups that
+      # may merge neighbouring cells. Widest sweep (4 values straddling 2.7).
+      molecular_distance_cutoff: [1.5, 2.0, 3.5, 4.0]
+      # ---- Tier 1: cell-flagging aggressiveness ----
+      # lrtest_nlog10P cutoff to flag putative mis-segmented cells. Default 5; LOWER flags more
+      # cells for re-segmentation (aggressive correction), HIGHER is conservative.
+      flagCell_lrtest_cutoff: [3, 8]
+      # ---- Tier 1: transcript high/low-score class boundary (SVM) ----
+      # Transcript-score cutoff separating high vs low score classes. Default -2; straddle it.
+      svmClass_score_cutoff: [-3, -1]
+      # ---- Tier 1: spatial-merge constraint ----
+      # Fractional (0-1) spatial constraint for accepting a group-merge event. Default 0.5;
+      # lower is stricter (fewer merges), higher is more permissive.
+      cutoff_spatialMerge: [0.3, 0.7]