fix(analysis): read R's NA as missing so parametric fits run

CHANGELOG.md

@@ -12,6 +12,27 @@ types, tests, infrastructure, and alignment.
 
 ## [Unreleased]
 
+### Fixed — every parametric fit returned `not_run`; CI failures are now readable (2026-09-26)
+
+- **R's `NA` is read as missing.** The estimator writes its fits with
+  `write.csv(..., na = "NA")` and read them back with CSV.jl's default
+  `missingstring = ""`. One `NA` in a numeric column made the whole column a string
+  column, `isfinite` threw, and every NB/logistic/Gaussian fit came back as `not_run`
+  with no statistics. The Julia side now reads `"NA"` as missing (`R_NA_STRINGS`).
+  Julia tests on `main` had been red since #60 for this reason.
+- **`glmGamPoi` reaches its by-name refusal.** The configuration lower-cases
+  `dispersion_method` but compared it against a table spelling `glmGamPoi`, so the name
+  was turned away as unknown and the issue #21 explanation was never shown. The compare
+  is now lower-case against lower-case, and the estimator looks its refusals up
+  case-insensitively. A test covers three spellings.
+- **Source-scan tests no longer trip on comments.** Comments quoting the removed
+  hash-derived p-value code were reworded so the "no placeholder statistics" scan stays
+  strict without flagging its own history.
+- **CI names the failing assertion.** The "Run tests" step posts one annotation per
+  `Test Failed` / `Error During Test` block and one with only the failing summary rows.
+  The old single annotation was cut to 4096 bytes by the checks API, which removed the
+  failing testset, and the job log is served from a host some environments cannot reach.
+
 ### Fixed — the ILR path stops substituting a basis it was not asked for (2026-09-25)
 
 - **Unimplemented ILR bases are refused rather than substituted.** The configuration

src/analysis/AnalysisConfig.jl

@@ -335,7 +335,10 @@ struct AdvancedConfig
         zero_policy_clean = lowercase(strip(zero_policy))
 
         # Dispersion method validation
-        if !(dispersion_method_clean in VALID_DISPERSION_METHODS)
+        # Lower-case against lower-case: the table spells `glmGamPoi` the way the package
+        # does, and comparing a lower-cased input against it refused the name at the door, so
+        # the estimator's by-name refusal (issue #21) was never reached.
+        if !(dispersion_method_clean in lowercase.(VALID_DISPERSION_METHODS))
             throw(ArgumentError("dispersion_method must be one of $(join(VALID_DISPERSION_METHODS, ", ")) — got '$dispersion_method_clean'. See context_help('advanced.dispersion_method')"))
         end
 

src/analysis/Execution.jl

@@ -1433,7 +1433,7 @@ end
 # a fit cannot be run.
 #
 # `Estimation.estimate_models` is the estimator. The block that used to stand here derived its
-# p-values from `hash(taxon_id)`, which is a function of the feature name and not of the data:
+# p-values from a hash of the taxon id, which is a function of the feature name and not of the data:
 # it has been removed rather than deprecated, and test/unit/test_estimation.jl asserts at the
 # source level that it does not come back.
 const ESTIMATION_R_WAIT_SECONDS = Ref(10.0)
@@ -1575,7 +1575,7 @@ function run_analysis(
     # ----------------------------------------------------------------------
     # Estimation — every number below comes from a fit that ran
     #
-    # This replaces a block that computed `p_val = 0.01 + (hash(taxon_id) % 100)/1000.0`
+    # This replaces a block that computed `p_val = 0.01 + (h % 100)/1000.0` with `h` the hash of the taxon id
     # and `padj = p_val * 1.5` and returned them as results. Those numbers were a
     # function of the feature NAME: reproducible, plausible-looking and meaningless.
     # See src/analysis/estimation.jl for what runs instead and

src/analysis/estimation.jl

@@ -9,7 +9,7 @@
 #
 # This module exists because the layer it replaces did not compute anything. The block
 # it replaces in Execution.run_analysis derived its p-values from
-# `p_val = 0.01 + (h % 100) / 1000.0` where `h = hash(taxon_id)` — a deterministic
+# `p_val = 0.01 + (h % 100) / 1000.0` where `h` was the hash of the taxon id — a deterministic
 # function of the feature NAME, carrying no information about the counts. The source
 # called it a stub; the caller received it as a result. A stub that is returned as a
 # result is a wrong answer, so it is gone rather than deprecated.
@@ -77,6 +77,28 @@ const REFUSED_DISPERSION = Dict{String,String}(
 
 const SUPPORTED_DISPERSION = ("parametric",)
 
+"""
+The strings R's `write.csv(..., na = "NA")` uses for a missing value. CSV.jl's default
+`missingstring` is only `""`, so without this a numeric column holding one `NA` is read as a
+string column and every later `isfinite` throws — which turned every fit into `not_run`.
+"""
+const R_NA_STRINGS = ["NA", ""]
+
+"""
+    _refused_dispersion(name) -> Union{Nothing,Pair{String,String}}
+
+Look a dispersion name up in `REFUSED_DISPERSION` without regard to case: the configuration
+layer lower-cases what it stores, and a `glmgampoi` has to meet the same refusal, with the
+same reason, as `glmGamPoi`.
+"""
+function _refused_dispersion(name::AbstractString)
+    key = lowercase(strip(name))
+    for (k, v) in REFUSED_DISPERSION
+        lowercase(k) == key && return k => v
+    end
+    return nothing
+end
+
 """
     ESTIMATE_SCALE
 
@@ -385,8 +407,9 @@ function estimate_models(config::AnalysisConfig.AnalysisConfig,
     # -- method preconditions (configuration errors: hard refusals) -------
     if method == AnalysisConfig.NB_GLM
         dispersion = lowercase(strip(config.advanced.dispersion_method))
-        haskey(REFUSED_DISPERSION, dispersion) &&
-            throw(ArgumentError("dispersion_method '$dispersion' is $(REFUSED_DISPERSION[dispersion]) See issue #21 and docs/statistics/method-conditions/parametric-fits.md."))
+        refused = _refused_dispersion(dispersion)
+        isnothing(refused) ||
+            throw(ArgumentError("dispersion_method '$(first(refused))' is $(last(refused)) See issue #21 and docs/statistics/method-conditions/parametric-fits.md."))
         dispersion in SUPPORTED_DISPERSION ||
             throw(ArgumentError("dispersion_method '$dispersion' is not one of $(join(SUPPORTED_DISPERSION, ", ")). Refusing."))
 
@@ -450,8 +473,8 @@ function estimate_models(config::AnalysisConfig.AnalysisConfig,
                 "mass"           => RCall.rcopy(String, RCall.reval("est_mass_version")),
                 "primary_levels" => RCall.rcopy(String, RCall.reval("est_primary_levels_text")),
             )
-            fit_rows[] = collect(CSV.File(fits_path))
-            coef_rows[] = collect(CSV.File(coefs_path))
+            fit_rows[] = collect(CSV.File(fits_path; missingstring = R_NA_STRINGS))
+            coef_rows[] = collect(CSV.File(coefs_path; missingstring = R_NA_STRINGS))
             RCall.reval("rm(list = ls(pattern = \"^est_\")); gc()")
             return nothing
         end

test/unit/test_estimation.jl

@@ -182,6 +182,21 @@
         @test occursin("glmGamPoi", err.msg)
         @test occursin("#21", err.msg)
 
+        # The configuration layer lower-cases what it stores; every spelling has to reach the
+        # same by-name refusal instead of being turned away as an unknown method at the door.
+        for spelling in ("glmgampoi", "GLMGAMPOI", " glmGamPoi ")
+            spelled = try
+                Estimation.estimate_models(nb_config(dispersion = spelling), counts_effect;
+                                           sample_metadata = meta_effect, offset = offsets_effect,
+                                           taxa_ids = taxa_effect)
+                nothing
+            catch e
+                e
+            end
+            @test spelled isa ArgumentError
+            @test occursin("glmGamPoi", spelled.msg) && occursin("#21", spelled.msg)
+        end
+
         # Counting without an offset would compare library sizes instead of groups.
         @test_throws ArgumentError Estimation.estimate_models(
             nb_config(), counts_effect; sample_metadata = meta_effect, taxa_ids = taxa_effect)