diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 1b1f80c5..7f743574 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -320,10 +320,45 @@ jobs:
           if-no-files-found: error
           retention-days: 30
 
+      # A failing assertion must not discard the coverage earned by every test
+      # that did run. Keep the report available for diagnosis, but only publish
+      # a complete passing run to Codecov. The test step's exit status still
+      # controls the job; these informational reports cannot make it green.
       - name: Process coverage
+        if: ${{ !cancelled() && hashFiles('**/*.cov') != '' }}
         uses: julia-actions/julia-processcoverage@03114f09f119417c3242a9fb6e0b722676aedf38  # v1.2.2
 
+      - name: Report descriptive-statistics coverage gaps
+        if: ${{ !cancelled() && hashFiles('lcov.info') != '' }}
+        shell: bash
+        run: |
+          set -euo pipefail
+          {
+            echo '## Descriptive-statistics coverage gaps'
+            echo
+            echo 'Informational only; the Julia test result remains the gate.'
+            echo
+            echo '| Source | Uncovered line |'
+            echo '| --- | ---: |'
+          } >> "$GITHUB_STEP_SUMMARY"
+          awk -F'[:,]' '
+            /^SF:src\/(canonical|server\/routes\/validated_statistics|analysis\/(derived_analyses|exact_summaries|high_precision|numeric_policy|method_catalogue))\.jl$/ {
+              source = substr($0, 4); next
+            }
+            source != "" && $1 == "DA" && $3 == 0 {
+              printf "::notice file=%s,line=%s::Julia coverage gap on this descriptive-statistics source line\n", source, $2
+              printf "| `%s` | %s |\n", source, $2 >> summary
+              gaps++
+            }
+            /^end_of_record$/ { source = "" }
+            END {
+              if (!gaps) print "| No uncovered lines in the descriptive-statistics sources | — |" >> summary
+              printf "Descriptive-statistics source lines without coverage: %d\n", gaps
+            }
+          ' summary="$GITHUB_STEP_SUMMARY" lcov.info
+
       - name: Retain raw Julia coverage
+        if: ${{ !cancelled() && hashFiles('lcov.info') != '' }}
         uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02  # v4.6.2
         with:
           name: julia-coverage
@@ -332,6 +367,7 @@ jobs:
           retention-days: 30
 
       - name: Upload coverage to Codecov
+        if: ${{ steps.julia_tests.outcome == 'success' }}
         uses: codecov/codecov-action@fb8b3582c8e4def4969c97caa2f19720cb33a72f  # v6.0.2
         with:
           files: lcov.info
diff --git a/docs/statistics/validation-evidence.md b/docs/statistics/validation-evidence.md
index c558c17f..1eaa7478 100644
--- a/docs/statistics/validation-evidence.md
+++ b/docs/statistics/validation-evidence.md
@@ -76,7 +76,17 @@ names colliding with projected feature labels, and subgroup names such as
 `A` and `AB`. Reserved integer names such as `total` are explicitly rejected
 when their count/metadata role is ambiguous. Subgroups use the same underscore-delimited prefix contract as
 the existing analysis routes. Tests check the persisted results and selected
-sample columns as well as the response.
+sample columns as well as the response. Quoted column identifiers are exercised
+through executed text/include/exclude predicates and checked against the exact
+selected counts; numeric filters on count columns remain explicitly refused
+before the shared floating-point SQL predicate can be used. BIGNUM columns
+cannot be silently omitted, and grouped runs persist and retrieve the same
+resolved group in their provenance.
+
+CI retains raw Julia coverage even when a test assertion fails, and lists
+uncovered descriptive-statistics source lines in the job summary and check
+annotations. These reports are diagnostic, not a replacement for a passing
+suite; only a passing test run publishes its coverage to Codecov.
 
 `frontend/src/components/AnalysisWorkspace.test.ts` renders the workspace and
 descriptive panel, toggles aggregation, and checks the submitted scope,
diff --git a/test/unit/test_derived_analyses.jl b/test/unit/test_derived_analyses.jl
index 16cd8ae2..94d477ed 100644
--- a/test/unit/test_derived_analyses.jl
+++ b/test/unit/test_derived_analyses.jl
@@ -25,6 +25,40 @@ using MetaManifold.NumericPolicy
     end
 end
 
+@testset "Derived-record I/O and storage limits fail explicitly" begin
+    DA = MetaManifold.DerivedAnalyses
+    mktempdir() do root
+        # Reading a directory raises an I/O error, independently of filesystem
+        # permissions (the test must also work when CI runs as root).
+        unreadable = joinpath(root, "not-a-file")
+        mkdir(unreadable)
+        error = try
+            DA._existing(unreadable, repeat("0", 64))
+            nothing
+        catch err
+            err
+        end
+        @test error isa DerivedAnalysisError
+        @test error.code == "conflict"
+        @test occursin("cannot read existing derived analysis", error.reason)
+
+        # The limit applies to the complete canonical record, not just payload.
+        # A payload at the limit necessarily makes the enclosing record exceed it.
+        directory = joinpath(root, "oversized")
+        error = try
+            persist_derived_analysis(directory;
+                source=Dict("run" => "r"), config=Dict("mode" => "exact_counts"),
+                result=Dict("payload" => repeat("x", MAX_DERIVED_BYTES)))
+            nothing
+        catch err
+            err
+        end
+        @test error isa DerivedAnalysisError
+        @test error.code == "resource_limit"
+        @test !isdir(directory) # refuse before allocating a publication directory
+    end
+end
+
 @testset "Immutable content-addressed derived analyses" begin
     mktempdir() do root
         directory = joinpath(root, "derived")
diff --git a/test/unit/test_validated_statistics.jl b/test/unit/test_validated_statistics.jl
index b2742beb..9724d117 100644
--- a/test/unit/test_validated_statistics.jl
+++ b/test/unit/test_validated_statistics.jl
@@ -198,11 +198,11 @@ function _vs_make_run(root::String)
             DBInterface.execute(con, "CREATE TABLE \"$(saved_name)\" AS SELECT * FROM merged")
         end
         DBInterface.execute(con, """CREATE TABLE merged_aliases (
-            SeqName VARCHAR, Genus VARCHAR, feature BIGINT,
+            SeqName VARCHAR, Genus VARCHAR, Pident DOUBLE, feature BIGINT,
             __VALIDATED_FEATURE BIGINT, __validated_feature_ BIGINT, "odd""sample" HUGEINT)""")
         DBInterface.execute(con, """INSERT INTO merged_aliases VALUES
-            ('seq2', 'G', 2, 20, 200, 9007199254740993),
-            ('seq1', 'G', 1, 10, 100, 1), ('seq3', 'H', 4, 40, 400, 3)""")
+            ('seq2', 'G', 99.0, 2, 20, 200, 9007199254740993),
+            ('seq1', 'G', 98.0, 1, 10, 100, 1), ('seq3', 'H', 97.0, 4, 40, 400, 3)""")
         # An unbounded integer is an integer the detector cannot name, so it
         # must be refused rather than silently dropped.
         DBInterface.execute(con, "CREATE TABLE merged_bignum (SeqName VARCHAR, Genus VARCHAR, s1 BIGINT, counted BIGNUM)")
@@ -265,7 +265,12 @@ end
             invalid_response = _vs_post(path, merge(request, Dict("table" => invalid)))
             @test invalid_response.status == 400
         end
-        @test _vs_post(path, request).status == 200 # invalid names did not mutate the source
+        # A safe name can still name a table that does not exist. Preserve the
+        # shared opener's 404 rather than treating its response as an analysis.
+        missing = _vs_post(path, merge(request, Dict("table" => "absent")))
+        @test missing.status == 404
+        @test JSON3.read(String(missing.body)).error == "table_not_found"
+        @test _vs_post(path, request).status == 200 # invalid/missing names did not mutate the source
         sequence_request = filter(pair -> first(pair) != "taxonomic_level", request)
         sequence_response = _vs_post(path, merge(sequence_request, Dict("time_limit_seconds" => 30)))
         @test sequence_response.status == 200
@@ -469,14 +474,33 @@ end
             response = _vs_post(path, merge(request, filters))
             @test response.status == 200
         end
-        # The quoted identifier itself also builds valid SQL. `min` on a sample
-        # column is refused by the exact-integer count-filter guard, which runs
-        # after the predicate is built, so reaching that refusal -- and not a
-        # 500 from malformed SQL -- is the regression under test.
-        quoted = _vs_post(path, merge(request,
-            Dict("colFilters" => Dict("odd\"sample" => Dict("min" => 1)))))
-        @test quoted.status == 422
-        @test occursin("count-filter", String(quoted.body))
+        # Execute text, include and exclude predicates against the quoted
+        # identifier itself. Each selects only seq2, so a skipped or malformed
+        # predicate cannot pass by returning the wrong rows.
+        for criterion in (Dict("text" => "9007199254740993"),
+                          Dict("include" => ["9007199254740993"]),
+                          Dict("exclude" => ["1", "3"]))
+            response = _vs_post(path, merge(request,
+                Dict("colFilters" => Dict("odd\"sample" => criterion))))
+            @test response.status == 200
+            result = JSON3.read(String(response.body))
+            @test collect(result.samples[1].features) == ["G"]
+            @test collect(result.samples[1].counts) == [2, 20, 200, "9007199254740993"]
+            @test result.samples[1].total == "9007199254741215"
+            loaded_response = _vs_get("/api/v1/studies/studyV/runs/run1/statistics/derived/$(result.analysis_id)")
+            @test loaded_response.status == 200
+            loaded = JSON3.read(String(loaded_response.body))
+            @test loaded.source.table == "merged_aliases"
+            @test loaded.result.samples == result.samples
+        end
+        # Numeric count filters are deliberately refused before SQL construction:
+        # the shared min/max predicate would cast exact counts to Float64.
+        for criterion in (Dict("min" => 1), Dict("max" => 9007199254740993))
+            quoted = _vs_post(path, merge(request,
+                Dict("colFilters" => Dict("odd\"sample" => criterion))))
+            @test quoted.status == 422
+            @test occursin("count-filter", String(quoted.body))
+        end
     end
 end