chore: update Gonum fork and require Go 1.27.1

.codex/environments/environment.toml

@@ -86,7 +86,7 @@ cd "$tmp"
 go mod init example.com/controlsys-consumer
 go mod edit -require=github.com/jamestjsp/controlsys@v0.0.0
 go mod edit -replace=github.com/jamestjsp/controlsys="$repo_path"
-go mod edit -replace=gonum.org/v1/gonum=github.com/jamestjsp/gonum@v0.17.3-fork
+go mod edit -replace=gonum.org/v1/gonum=github.com/jamestjsp/gonum@v0.19.0-fork
 cat > controlsys_consumer_test.go <<'EOF'
 package controlsysconsumer
 

.github/workflows/ci.yml

@@ -1,7 +1,7 @@
 name: CI
 on: [push, pull_request]
 env:
-  GO_VERSION: '1.26.3'
+  GO_VERSION: '1.27.1'
 jobs:
   test:
     runs-on: ubuntu-latest
@@ -28,7 +28,7 @@ jobs:
           go mod init example.com/controlsys-consumer
           go mod edit -require=github.com/jamestjsp/controlsys@v0.0.0
           go mod edit -replace=github.com/jamestjsp/controlsys="$GITHUB_WORKSPACE"
-          go mod edit -replace=gonum.org/v1/gonum=github.com/jamestjsp/gonum@v0.17.3-fork
+          go mod edit -replace=gonum.org/v1/gonum=github.com/jamestjsp/gonum@v0.19.0-fork
           cat > controlsys_consumer_test.go <<'EOF'
           package controlsysconsumer
 

README.md

@@ -12,7 +12,7 @@ go get github.com/jamestjsp/controlsys
 
 > **Note:** This package depends on a [gonum fork](https://github.com/jamestjsp/gonum) for additional LAPACK routines. Because `replace` directives do not propagate to downstream modules, applications that import `controlsys` must add this to their own `go.mod`:
 > ```
-> replace gonum.org/v1/gonum => github.com/jamestjsp/gonum v0.17.3-fork
+> replace gonum.org/v1/gonum => github.com/jamestjsp/gonum v0.19.0-fork
 > ```
 
 ## Production Readiness

docs/benchmarks/gonum-v0.19.0/README.md

@@ -0,0 +1,110 @@
+# Gonum v0.19.0-fork upgrade: performance evidence
+
+For the updated PR requiring Go 1.27.1, see the [full follow-up](go1.27.1-full/README.md).
+It substantially reduces the large regressions below, but confirms residual
+slowdowns up to 6.17%. This original Go 1.26.4 report remains historical evidence.
+
+Measured on 2026-09-12 for PR #171. The upgrade has mixed performance on this
+machine. The complete 236-case timing geomean improves 0.75%, but reproducible
+slowdowns remain. This is not a regression-free performance approval. Assess the
+tradeoff or repair the affected backend paths before merging and publishing the
+prepared v1.5.1 release.
+
+## Comparison
+
+- Baseline: `fdda4f150d13b34a969de7faed6ae896e827ce94`, Gonum `v0.17.7-fork`.
+- Candidate: `54d5d6b119c8b4351db0cd0bc0ebb4122bde94d4`, Gonum `v0.19.0-fork`.
+- The baseline was the exact PR parent and freshly fetched `origin/main`.
+- Apple M1 Pro, darwin/arm64, Go 1.26.4, `GOMAXPROCS=8`.
+- Default Gonum Go BLAS and LAPACK implementations; no backend overrides,
+  custom build tags, GOFLAGS, or PGO profiles. Native compilation, no emulation.
+- Same unchanged benchmark fixtures on both revisions. Both test binaries were
+  built before timing. No agent builds, tests, or profiling ran concurrently.
+  This was a normal desktop host, without CPU affinity or power-state control.
+- Ten samples per case per revision, alternating baseline/candidate and
+  candidate/baseline order. Full suite: 100ms per benchmark; focused checks:
+  300ms. All benchmark processes exited successfully with matching case sets.
+- `benchstat`: `golang.org/x/perf v0.0.0-20260312031701-16a31bc5fbd0`.
+
+## Results
+
+The longer follow-up covered the 14 full-sweep slowdowns of at least 3%, plus
+DC motor simulation, regulator construction, and Lyapunov allocation behavior.
+One nested benchmark needed a separate slash-separated selection. Medians below
+come from the longer samples; all listed changes have p < 0.001, except the
+10,000-point SISO FRD case (p=0.009). Full tables include confidence intervals.
+
+| Benchmark | Baseline | Candidate | Time change |
+| --- | ---: | ---: | ---: |
+| MatLog_N50 | 330.3 us | 372.0 us | +12.62% |
+| D2C_ZOH_N50 | 413.3 us | 454.3 us | +9.92% |
+| Stabsep_N100 | 8.904 ms | 9.455 ms | +6.18% |
+| SystemFRD_SISO_10000 | 220.2 us | 233.7 us | +6.15% |
+| Reduce | 45.94 us | 48.43 us | +5.42% |
+| Modsep_N50 | 1.429 ms | 1.504 ms | +5.29% |
+| Stabsep_N50 | 1.385 ms | 1.458 ms | +5.27% |
+| DiscretizeZOH | 21.48 us | 22.58 us | +5.15% |
+| MatLog_N20 | 49.37 us | 51.78 us | +4.88% |
+| ControllabilityStaircase | 36.36 us | 37.94 us | +4.36% |
+| D2C_ZOH_N20 | 61.47 us | 63.94 us | +4.02% |
+| Canon_Modal_N50 | 640.5 us | 660.9 us | +3.19% |
+| Stabsep_N10 | 32.70 us | 33.47 us | +2.37% |
+| Simulate_DCMotor | 46.69 us | 39.72 us | -14.93% |
+| Reg_N100_M5_P5 | 79.72 us | 60.92 us | -23.58% |
+
+The `FrequencySweepKernels/N4_W2/TransferFunction` slowdown did not reproduce
+with longer samples (p=0.447). Lyap_N100 timing was inconclusive (p=0.353).
+Absence of statistical significance does not establish equivalence. The full
+sweep contains additional smaller changes; see its complete table rather than
+using this selected follow-up as an aggregate performance score. Multiple case
+comparisons can produce false positives, especially for small effects.
+
+Allocations also vary: DC motor simulation drops from 45 to 11 allocs/op and
+regulator construction from 55 to 25. The focused SISO FRD case rises from 60 to
+61, and Stabsep_N100 from 481 to 482. Lyap_N100's full-sweep increase from 363 to
+364 did not reproduce in the focused run (both medians 363), although its bytes
+per operation increased 0.08%. Tables retain all memory measurements.
+
+These data establish behavior only on this host at eight workers. They do not
+establish performance on amd64, other worker counts, or arbitrary models, and do
+not attribute a specific backend change as the cause. Correctness tests and race
+tests passed separately; performance measurements do not replace those tests.
+
+## Evidence and reproduction
+
+- `full-benchstat.txt`: all 236 cases, timing and allocations.
+- `focused-benchstat.txt`: 16 longer follow-up cases.
+- `subcase-benchstat.txt`: the nested transfer-function follow-up.
+- Matching `*-baseline.txt.gz` and `*-candidate.txt.gz`: complete raw samples.
+- `metadata.json`: full-sweep binary hashes, flags, environment, and run order.
+
+From the repository, build isolated copies of the exact revisions, then run the
+existing Go benchmarks. Prebuilding is equivalent to `go test -run '^$'
+-bench=. -benchmem -benchtime=100ms -count=1`, with compilation outside timing.
+
+```bash
+set -eu
+bench_dir=$(mktemp -d)
+mkdir "$bench_dir/baseline" "$bench_dir/candidate"
+git archive fdda4f150d13b34a969de7faed6ae896e827ce94 | tar -x -C "$bench_dir/baseline"
+git archive 54d5d6b119c8b4351db0cd0bc0ebb4122bde94d4 | tar -x -C "$bench_dir/candidate"
+(cd "$bench_dir/baseline" && go test -c -o "$bench_dir/baseline.test")
+(cd "$bench_dir/candidate" && go test -c -o "$bench_dir/candidate.test")
+export GOMAXPROCS=8
+for round in $(seq 1 10); do
+  order='baseline candidate'
+  if [ $((round % 2)) -eq 0 ]; then order='candidate baseline'; fi
+  for label in $order; do
+    "$bench_dir/$label.test" -test.run='^$' -test.bench=. \
+      -test.benchmem -test.benchtime=100ms -test.count=1 \
+      >> "$bench_dir/$label.txt"
+  done
+done
+benchstat "$bench_dir/baseline.txt" "$bench_dir/candidate.txt"
+```
+
+For focused reproduction, use 300ms and select the top-level names in
+`focused-benchstat.txt` with `^Benchmark(Name1|Name2|...)$`. For the nested case,
+use `^BenchmarkFrequencySweepKernels$/^N4_W2$/^TransferFunction$` separately.
+To reanalyze retained data, decompress each pair with `gzip -dc` and pass the
+resulting files to the recorded benchstat version.

docs/benchmarks/gonum-v0.19.0/focused-benchstat.txt

@@ -0,0 +1,64 @@
+goos: darwin
+goarch: arm64
+pkg: github.com/jamestjsp/controlsys
+cpu: Apple M1 Pro
+                           │ baseline.txt │            candidate.txt            │
+                           │    sec/op    │   sec/op     vs base                │
+Reduce-8                      45.94µ ± 1%   48.43µ ± 7%   +5.42% (p=0.000 n=10)
+DiscretizeZOH-8               21.48µ ± 1%   22.58µ ± 6%   +5.15% (p=0.000 n=10)
+ControllabilityStaircase-8    36.36µ ± 2%   37.94µ ± 3%   +4.36% (p=0.000 n=10)
+Simulate_DCMotor-8            46.69µ ± 1%   39.72µ ± 2%  -14.93% (p=0.000 n=10)
+Lyap_N100-8                   4.052m ± 1%   4.048m ± 0%        ~ (p=0.353 n=10)
+Reg_N100_M5_P5-8              79.72µ ± 7%   60.92µ ± 2%  -23.58% (p=0.000 n=10)
+D2C_ZOH_N20-8                 61.47µ ± 1%   63.94µ ± 1%   +4.02% (p=0.000 n=10)
+D2C_ZOH_N50-8                 413.3µ ± 1%   454.3µ ± 1%   +9.92% (p=0.000 n=10)
+MatLog_N20-8                  49.37µ ± 1%   51.78µ ± 1%   +4.88% (p=0.000 n=10)
+MatLog_N50-8                  330.3µ ± 1%   372.0µ ± 3%  +12.62% (p=0.000 n=10)
+SystemFRD_SISO_10000-8        220.2µ ± 6%   233.7µ ± 3%   +6.15% (p=0.009 n=10)
+Canon_Modal_N50-8             640.5µ ± 1%   660.9µ ± 2%   +3.19% (p=0.000 n=10)
+Stabsep_N10-8                 32.70µ ± 1%   33.47µ ± 1%   +2.37% (p=0.000 n=10)
+Stabsep_N50-8                 1.385m ± 1%   1.458m ± 3%   +5.27% (p=0.000 n=10)
+Stabsep_N100-8                8.904m ± 2%   9.455m ± 4%   +6.18% (p=0.000 n=10)
+Modsep_N50-8                  1.429m ± 3%   1.504m ± 3%   +5.29% (p=0.000 n=10)
+geomean                       213.2µ        217.1µ        +1.84%
+
+                           │ baseline.txt │            candidate.txt             │
+                           │     B/op     │     B/op      vs base                │
+Reduce-8                     52.43Ki ± 0%   52.44Ki ± 0%   +0.02% (p=0.002 n=10)
+DiscretizeZOH-8              21.67Ki ± 0%   21.67Ki ± 0%        ~ (p=0.696 n=10)
+ControllabilityStaircase-8   22.16Ki ± 0%   22.15Ki ± 0%        ~ (p=0.126 n=10)
+Simulate_DCMotor-8           19.67Ki ± 0%   16.28Ki ± 0%  -17.22% (p=0.000 n=10)
+Lyap_N100-8                  373.0Ki ± 0%   373.3Ki ± 0%   +0.08% (p=0.000 n=10)
+Reg_N100_M5_P5-8             424.2Ki ± 0%   421.4Ki ± 0%   -0.66% (p=0.000 n=10)
+D2C_ZOH_N20-8                77.60Ki ± 1%   77.40Ki ± 0%   -0.25% (p=0.035 n=10)
+D2C_ZOH_N50-8                463.5Ki ± 0%   463.2Ki ± 0%        ~ (p=0.218 n=10)
+MatLog_N20-8                 64.54Ki ± 0%   64.50Ki ± 0%        ~ (p=0.954 n=10)
+MatLog_N50-8                 391.2Ki ± 0%   391.8Ki ± 0%   +0.16% (p=0.007 n=10)
+SystemFRD_SISO_10000-8       922.6Ki ± 0%   922.6Ki ± 0%   +0.01% (p=0.000 n=10)
+Canon_Modal_N50-8            209.7Ki ± 0%   211.0Ki ± 0%   +0.58% (p=0.001 n=10)
+Stabsep_N10-8                29.33Ki ± 0%   29.32Ki ± 0%        ~ (p=0.288 n=10)
+Stabsep_N50-8                635.1Ki ± 0%   635.4Ki ± 0%        ~ (p=0.796 n=10)
+Stabsep_N100-8               2.528Mi ± 0%   2.532Mi ± 0%        ~ (p=0.218 n=10)
+Modsep_N50-8                 640.9Ki ± 0%   640.3Ki ± 0%        ~ (p=0.218 n=10)
+geomean                      168.6Ki        166.6Ki        -1.18%
+
+                           │ baseline.txt │            candidate.txt             │
+                           │  allocs/op   │ allocs/op   vs base                  │
+Reduce-8                       170.0 ± 0%   170.0 ± 0%        ~ (p=1.000 n=10) ¹
+DiscretizeZOH-8                28.00 ± 0%   28.00 ± 0%        ~ (p=1.000 n=10) ¹
+ControllabilityStaircase-8     109.0 ± 0%   109.0 ± 0%        ~ (p=1.000 n=10) ¹
+Simulate_DCMotor-8             45.00 ± 0%   11.00 ± 0%  -75.56% (p=0.000 n=10)
+Lyap_N100-8                    363.0 ± 0%   363.0 ± 0%        ~ (p=0.474 n=10)
+Reg_N100_M5_P5-8               55.00 ± 0%   25.00 ± 0%  -54.55% (p=0.000 n=10)
+D2C_ZOH_N20-8                  50.00 ± 0%   50.00 ± 0%        ~ (p=1.000 n=10) ¹
+D2C_ZOH_N50-8                  53.00 ± 0%   53.00 ± 0%        ~ (p=1.000 n=10) ¹
+MatLog_N20-8                   30.00 ± 0%   30.00 ± 0%        ~ (p=1.000 n=10) ¹
+MatLog_N50-8                   32.00 ± 0%   32.00 ± 0%        ~ (p=1.000 n=10) ¹
+SystemFRD_SISO_10000-8         60.00 ± 2%   61.00 ± 0%   +1.67% (p=0.001 n=10)
+Canon_Modal_N50-8              112.0 ± 0%   112.0 ± 0%        ~ (p=1.000 n=10) ¹
+Stabsep_N10-8                  79.00 ± 0%   79.00 ± 0%        ~ (p=1.000 n=10) ¹
+Stabsep_N50-8                  156.0 ± 1%   156.0 ± 0%        ~ (p=0.211 n=10)
+Stabsep_N100-8                 481.0 ± 0%   482.0 ± 0%   +0.21% (p=0.002 n=10)
+Modsep_N50-8                   143.0 ± 1%   143.0 ± 0%        ~ (p=0.087 n=10)
+geomean                        84.50        73.74       -12.73%
+¹ all samples are equal