Gradient classes: derive from parameters_fresh, optional override arg #45

Problem

R/lnk_pipeline_prepare.R carries two hardcoded lists that encode bcfishpass's gradient-classification scheme directly in pipeline code. Both belong at the config/derivation layer.

Hardcode 1 — gradient class set in .lnk_pipeline_prep_gradient:

fresh::frs_break_find(conn, paste0(schema, ".streams_blk"),
  attribute = "gradient",
  classes = c("1500" = 0.15, "2000" = 0.20,
              "2500" = 0.25, "3000" = 0.30),
  to = paste0(schema, ".gradient_barriers_raw"))

These four classes are bcfishpass's 0.05-step bins spanning the union of per-species access_gradient_max values in parameters_fresh.csv (0.15, 0.20, 0.25) plus one bcfishpass-convention extra bin at 0.30. Any variant using different per-species access tolerances or a different binning scheme is blocked by this hardcode.

Hardcode 2 — per-species class filters in .lnk_pipeline_prep_minimal:

models <- list(
  bt              = c(2500, 3000),
  ch_cm_co_pk_sk  = c(1500, 2000, 2500, 3000),
  st              = c(2000, 2500, 3000),
  wct             = c(2000, 2500, 3000)
)

These encode "for each species, which gradient classes count as access barriers" — which is fully derivable from each species's access_gradient_max: a class value g is a barrier for species s when g >= s$access_gradient_max. The in-code TODO on this one already flagged it for follow-up.

What each encodes, and where it should live

• The class set is a modelling choice — what gradient bins does the pipeline detect? bcfishpass picks 0.05-step bins from 0.15 to 0.30. Other variants could choose finer or coarser granularity. Decision belongs at the caller / config level.
• The per-species filter is derived from per-species access tolerances — no independent decision.

Design

API

lnk_pipeline_prepare accepts an optional classes override:

lnk_pipeline_prepare <- function(conn, aoi, cfg, schema,
                                 observations = "bcfishobs.observations",
                                 classes = NULL) {
  if (is.null(classes)) {
    classes <- .lnk_derive_gradient_classes(cfg$parameters_fresh)
  }
  # pass classes through to .lnk_pipeline_prep_gradient and
  # .lnk_pipeline_prep_minimal
}

Default derivation

Internal @noRd helper:

.lnk_derive_gradient_classes <- function(params_fresh) {
  g <- sort(unique(params_fresh$access_gradient_max))
  g <- g[!is.na(g) & g > 0]
  setNames(g, sprintf("%04d", round(g * 10000)))
}

For the bundled bcfishpass parameters_fresh.csv, this produces c("1500" = 0.15, "2000" = 0.20, "2500" = 0.25). Missing vs the current hardcode: the "3000" bin — bcfishpass convention, no functional impact on access decisions (every position ≥ 0.30 is already ≥ 0.25 and labels "2500", still blocks every species that considers "2500" a barrier).

Reproduction parity

bit-identical rollup on bcfishpass variant requires keeping the bcfishpass class labels exactly, so data-raw/compare_bcfishpass_wsg.R passes the explicit set:

link::lnk_pipeline_prepare(conn, aoi, cfg, schema,
  classes = c("1500" = 0.15, "2000" = 0.20,
              "2500" = 0.25, "3000" = 0.30))

Per-species filter derivation

.lnk_pipeline_prep_minimal derives at runtime, no hardcoded models list:

for (sp in cfg$species) {
  sp_amax <- cfg$parameters_fresh$access_gradient_max[
    cfg$parameters_fresh$species_code == sp]
  if (length(sp_amax) == 0 || is.na(sp_amax) || sp_amax <= 0) next
  classes_for_sp <- names(classes)[classes >= sp_amax]
  # build barriers_<species> table for this class set
  # call fresh::frs_barriers_minimal
}

Note: this switches from bcfishpass's four per-model groupings (bt / ch_cm_co_pk_sk / st / wct) to one per species. Union into gradient_barriers_minimal must still produce the same set of positions for bit-identical output — species in the same bcfishpass group share access_gradient_max values and therefore the same class filter, so union is equivalent.

Scope

1. Add internal helper .lnk_derive_gradient_classes() in R/utils.R or R/lnk_pipeline_prepare.R.
2. Extend lnk_pipeline_prepare() signature with classes = NULL; derive when NULL.
3. .lnk_pipeline_prep_gradient accepts the resolved classes and uses them.
4. .lnk_pipeline_prep_minimal derives per-species class filters from access_gradient_max, removes hardcoded models list.
5. data-raw/compare_bcfishpass_wsg.R passes the explicit bcfishpass class set.
6. Unit tests:
   • Derivation produces expected classes + names from the bcfishpass parameters_fresh.csv
   • Per-species filter logic picks the right classes for each species
   • Derivation handles NA / zero / missing column gracefully
7. Rerun tar_make() on all four WSGs. Rollup must be bit-identical to pre-change (passing the explicit class set preserves bcfishpass labels).
8. Short note in research/bcfishpass_comparison.md documenting that the class set is now config-passed rather than hardcoded.

Out of scope

• Moving classes into config.yaml$pipeline$gradient_classes. Deferred until we have a second variant that forces a different set.
• Scanning other columns in parameters_fresh (cluster_bridge_gradient, rear_gradient_max, etc.) as additional class sources. cluster_bridge_gradient is a continuous segment-level threshold with no role in class detection. Rule-threshold (rear_gradient_max, spawn_gradient_max) segmentation is a different pipeline concept — break positions at rule thresholds to reduce near-threshold averaging — and should be its own issue.

Risks

• Label mismatch on bcfishpass variant if the caller forgets to pass the explicit class set. Mitigation: compare_bcfishpass_wsg passes it mandatorily in this PR. A future config.yaml-driven path removes this footgun by making the list declarative in the bundle.
• Per-species union not equivalent to per-model union if access_gradient_max values in parameters_fresh don't actually match bcfishpass's grouping. Verify by asserting the post-change gradient_barriers_minimal row count equals pre-change row count per WSG.
• Species with NA or zero access_gradient_max (e.g. lake-only species with no stream access requirement) produce no filter. Derivation skips them — they contribute nothing to barrier detection. Test the skip path explicitly.
• Reproducibility must remain bit-identical across runs after the change.

Execution

• Branch from main. /planning-init + PWF.
• /code-check on staged diff before every commit.
• PR with SRED tag (Relates to NewGraphEnvironment/sred-2025-2026#24).

Cross-references

• Companion cleanup to Wire barriers_definite_control into lnk_barrier_overrides #44 (wire barriers_definite_control into lnk_barrier_overrides). Issues are parallel-safe — no ordering constraint. Both target R/lnk_pipeline_prepare.R but different sub-helpers.
• Follow-up issue (to file later): rule-threshold break source — break segments at rear_gradient_max / spawn_gradient_max crossings so classification near thresholds is deterministic. Different concept than this issue; different verification path.

Versions at start

• fresh: 0.14.0
• link: main (0.5.0, target 0.6.0)
• bcfishpass: ea3c5d8
• fwapg: Docker (FWA 20240830)