From df82c9e118be343ac017d2206361ccf87daaacee Mon Sep 17 00:00:00 2001
From: logan-nc <6198372+logan-nc@users.noreply.github.com>
Date: Fri, 14 Aug 2026 16:28:20 -0400
Subject: [PATCH 1/9] GPEC - NEW FEATURE - Record per-stage runtimes in gpec.h5
 under Info/Runtimes

main_from_inputs already timed every pipeline stage but only logged the
durations. Persist them under Info/Runtimes/<stage> (Float64 seconds, annotated
with long_name/units) so an output file answers "where did the time go" on its
own. Only the stages that ran are recorded; timings are informational, never a
regression quantity.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 docs/development/hdf5-conventions.md   |  2 +-
 src/GeneralizedPerturbedEquilibrium.jl | 66 ++++++++++++++++++++++----
 test/runtests_fullruns.jl              |  3 ++
 3 files changed, 62 insertions(+), 9 deletions(-)

diff --git a/docs/development/hdf5-conventions.md b/docs/development/hdf5-conventions.md
index b24610d2a..1048293e4 100644
--- a/docs/development/hdf5-conventions.md
+++ b/docs/development/hdf5-conventions.md
@@ -27,7 +27,7 @@ Top level (10 groups):
 
 | Group | Contents |
 |---|---|
-| `Info/` | Run metadata: `git_version`, mode-number ranges (`mpert`, `mlow`, …, `mn_index`), `psilim`, `qlim` |
+| `Info/` | Run metadata: `git_version`, mode-number ranges (`mpert`, `mlow`, …, `mn_index`), `psilim`, `qlim`, `Runtimes/` (per-stage wall-clock seconds) |
 | `Input/` | Rerun snapshot: `gpec_toml_raw`, `RawInputs/{Equilibrium, ForcingTerms, Coils/<name>}` |
 | `Equilibrium/` | Scalars (β, q₀, q95, …) plus `Profiles/` (1-D: xs, 2piF, mu0p, dVdpsi, q) and `Geometry/` (2-D: rcoords, offset, nu, jac) |
 | `ForceFreeStates/` | `Solutions/ForwardIntegration/` (u-solutions), `Solutions/GalerkinIntegration/` (`Solution/`, `Match/`, `msing`), `EulerLagrangeMatrices/{Ideal,Kinetic}`, `FreeBoundaryStability/`, `EdgeScan/` |
diff --git a/src/GeneralizedPerturbedEquilibrium.jl b/src/GeneralizedPerturbedEquilibrium.jl
index 656316b75..cbd5528f2 100755
--- a/src/GeneralizedPerturbedEquilibrium.jl
+++ b/src/GeneralizedPerturbedEquilibrium.jl
@@ -67,6 +67,7 @@ const H5_RAW_EQUILIBRIUM = "Input/RawInputs/Equilibrium"
 const H5_RAW_FORCING = "Input/RawInputs/ForcingTerms"
 const H5_RAW_COILS = "Input/RawInputs/Coils"
 const H5_GIT_VERSION = "Info/git_version"
+const H5_RUNTIMES = "Info/Runtimes"
 
 include("HDF5Schema.jl")
 include("Rerun.jl")
@@ -181,6 +182,8 @@ function main_from_inputs(
     preloaded_coil_sets::Union{Nothing,Vector{ForcingTerms.CoilSet}}=nothing
 )
     total_start = time()
+    # Per-stage wall-clock seconds, written to Info/Runtimes at the end of the run.
+    runtimes = Vector{Pair{String,Float64}}()
 
     # ----------------------------------------------------------------
     # Equilibrium
@@ -279,7 +282,9 @@ function main_from_inputs(
         @info "Two-pass psi grid: $(length(psi_nodes)) knots, $(length(mandatory)) rational surfaces pinned (n=$nstring)"
     end
 
-    @info "Equilibrium construction completed in $(@sprintf("%.3f", time() - equil_start)) s"
+    equil_dt = time() - equil_start
+    push!(runtimes, "equilibrium" => equil_dt)
+    @info "Equilibrium construction completed in $(@sprintf("%.3f", equil_dt)) s"
 
     # Early exit if user only requested equilibrium setup
     if equil.config.force_termination
@@ -498,7 +503,9 @@ function main_from_inputs(
     if ctrl.gal_flag
         gal_start = time()
         gal_data = galerkin_solve(ctrl, equil, ffit, intr; vac_data=ctrl.vac_flag ? vac_data : nothing)
-        @info "Galerkin solve completed in $(@sprintf("%.3f", time() - gal_start)) s"
+        gal_dt = time() - gal_start
+        push!(runtimes, "galerkin" => gal_dt)
+        @info "Galerkin solve completed in $(@sprintf("%.3f", gal_dt)) s"
     end
 
     if ctrl.write_outputs_to_HDF5
@@ -518,7 +525,9 @@ function main_from_inputs(
         @info "Results written to $(ctrl.HDF5_filename)"
     end
 
-    @info "Force-Free States completed in $(@sprintf("%.3f", time() - ffs_start)) s"
+    ffs_dt = time() - ffs_start
+    push!(runtimes, "force_free_states" => ffs_dt)
+    @info "Force-Free States completed in $(@sprintf("%.3f", ffs_dt)) s"
 
     # SLAYER tearing-mode analysis stage. Needs only equil + intr, so it runs in
     # both the force_termination=true path and the full pipeline. `pe_file` is the
@@ -536,7 +545,9 @@ function main_from_inputs(
             slayer_start = time()
             result = Runner.run_slayer(equil, intr, slayer_ctrl;
                 dir_path=intr.dir_path)
-            @info "SLAYER completed in $(@sprintf("%.3f", time() - slayer_start)) s"
+            slayer_dt = time() - slayer_start
+            push!(runtimes, "slayer" => slayer_dt)
+            @info "SLAYER completed in $(@sprintf("%.3f", slayer_dt)) s"
             h5_filename = pe_file === nothing ? ctrl.HDF5_filename : pe_file
             h5_path = joinpath(intr.dir_path, h5_filename)
             # Append the Tearing/ group; create the file if no prior stage wrote
@@ -557,7 +568,10 @@ function main_from_inputs(
     # Early exit if user only requested force-free states (SLAYER still runs).
     if ctrl.force_termination
         slayer_result = _run_slayer_stage(nothing)
-        @info "\n$_BANNER\n  GPEC completed successfully in $(@sprintf("%.3f", time() - total_start)) s\n$_BANNER"
+        total_dt = time() - total_start
+        push!(runtimes, "total" => total_dt)
+        _write_runtimes!(joinpath(intr.dir_path, ctrl.HDF5_filename), runtimes)
+        @info "\n$_BANNER\n  GPEC completed successfully in $(@sprintf("%.3f", total_dt)) s\n$_BANNER"
         return (ctrl=ctrl, equil=equil, intr=intr, ffit=ffit, odet=odet,
             vac_data=ctrl.vac_flag ? vac_data : nothing,
             slayer=slayer_result)
@@ -642,7 +656,9 @@ function main_from_inputs(
         end
     end
 
-    @info "Perturbed Equilibrium completed in $(@sprintf("%.3f", time() - pe_start)) s"
+    pe_dt = time() - pe_start
+    push!(runtimes, "perturbed_equilibrium" => pe_dt)
+    @info "Perturbed Equilibrium completed in $(@sprintf("%.3f", pe_dt)) s"
 
     # ----------------------------------------------------------------
     # KineticForces (Neoclassical Toroidal Viscosity)
@@ -670,7 +686,9 @@ function main_from_inputs(
             end
         end
 
-        @info "KineticForces completed in $(@sprintf("%.3f", time() - kf_start)) s"
+        kf_dt = time() - kf_start
+        push!(runtimes, "kinetic_forces" => kf_dt)
+        @info "KineticForces completed in $(@sprintf("%.3f", kf_dt)) s"
     end
 
     # ----------------------------------------------------------------
@@ -688,7 +706,10 @@ function main_from_inputs(
     # ----------------------------------------------------------------
     # Done
     # ----------------------------------------------------------------
-    @info "\n$_BANNER\n  GPEC completed successfully in $(@sprintf("%.3f", time() - total_start)) s\n$_BANNER"
+    total_dt = time() - total_start
+    push!(runtimes, "total" => total_dt)
+    _write_runtimes!(joinpath(intr.dir_path, ctrl.HDF5_filename), runtimes)
+    @info "\n$_BANNER\n  GPEC completed successfully in $(@sprintf("%.3f", total_dt)) s\n$_BANNER"
 
     # TODO: Do not allow perturbed equilibrium calculations if zero crossings are found
 
@@ -1017,6 +1038,35 @@ function _write_coil_snapshot!(h5_path::String, coil_sets::Vector{ForcingTerms.C
     return nothing
 end
 
+const _RUNTIME_LONG_NAMES = Dict(
+    "equilibrium" => "Wall-clock time of the Equilibrium construction stage",
+    "force_free_states" => "Wall-clock time of the ForceFreeStates stability stage",
+    "galerkin" => "Wall-clock time of the Galerkin outer-region solve",
+    "slayer" => "Wall-clock time of the SLAYER tearing-mode stage",
+    "perturbed_equilibrium" => "Wall-clock time of the PerturbedEquilibrium stage",
+    "kinetic_forces" => "Wall-clock time of the KineticForces (NTV) stage",
+    "total" => "Wall-clock time of the full GPEC run")
+
+"""
+    _write_runtimes!(h5_path::String, runtimes)
+
+Write the per-stage wall-clock seconds collected during a run into `Info/Runtimes/` of an
+existing gpec.h5 file. `runtimes` iterates `stage => seconds` pairs; only the stages that ran
+are recorded. These timings are informational only — machine- and load-dependent, never a
+regression quantity.
+"""
+function _write_runtimes!(h5_path::String, runtimes)
+    isfile(h5_path) || return nothing
+    h5open(h5_path, "r+") do out_h5
+        haskey(out_h5, H5_RUNTIMES) && HDF5.delete_object(out_h5, H5_RUNTIMES)
+        for (stage, dt) in runtimes
+            out_h5["$H5_RUNTIMES/$stage"] = dt
+        end
+        Utilities.HDF5Annotations.annotate!(out_h5, ["$H5_RUNTIMES/$stage" => (; long_name=_RUNTIME_LONG_NAMES[stage], units="s") for (stage, _) in runtimes])
+    end
+    return nothing
+end
+
 """
     write_imas(dd, result)
 
diff --git a/test/runtests_fullruns.jl b/test/runtests_fullruns.jl
index ce6470385..cb0adc19d 100644
--- a/test/runtests_fullruns.jl
+++ b/test/runtests_fullruns.jl
@@ -28,6 +28,9 @@ using HDF5
             et = read(h5["ForceFreeStates/FreeBoundaryStability/eigenmode_energies"])
             @test isfinite(real(et[1]))
             @test real(et[1]) > 0
+            # Per-stage wall-clock records (informational, not regression quantities).
+            @test haskey(h5, "Info/Runtimes/total") && read(h5["Info/Runtimes/total"]) > 0
+            @test haskey(h5, "Info/Runtimes/force_free_states")
         end
         rm(joinpath(ex3, "gpec.h5"); force=true)
         true

From e695e890e4a7564f1dd3e05d4e7f2b42365a7c64 Mon Sep 17 00:00:00 2001
From: logan-nc <6198372+logan-nc@users.noreply.github.com>
Date: Fri, 14 Aug 2026 16:29:45 -0400
Subject: [PATCH 2/9] BENCHMARKS - IMPROVEMENT - Report per-stage runtime
 deltas from Info/Runtimes

Collect Info/Runtimes inside the warm-run loop (each run overwrites gpec.h5),
average over the stages present in every run, and print a per-stage delta table
alongside the existing totals. Refs predating the record yield an empty dict and
the table is skipped silently, so old-vs-new comparisons still work.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 benchmarks/benchmark_git_branches.jl | 58 ++++++++++++++++++++++++++--
 1 file changed, 55 insertions(+), 3 deletions(-)

diff --git a/benchmarks/benchmark_git_branches.jl b/benchmarks/benchmark_git_branches.jl
index c8e017bc2..63367fb0e 100755
--- a/benchmarks/benchmark_git_branches.jl
+++ b/benchmarks/benchmark_git_branches.jl
@@ -43,6 +43,7 @@ For each branch/commit, reports:
 - Eigenmode energy (et[1]) from gpec.h5
 - Integration steps from gpec.h5
 - Runtime (averaged over warm runs)
+- Per-stage runtime breakdown from `Info/Runtimes` (averaged over warm runs, when both refs record it)
 - Git commit hash
 
 # Example Output
@@ -67,6 +68,12 @@ Comparison:
   Eigenmode Δ:    +0.0067 (+0.4%)
   Steps Δ:        -775 (-85.1%)
   Runtime Δ:      -2.5 s (-1.7% faster)
+
+Per-stage runtime (avg over warm runs):
+  Stage                    Branch 1   Branch 2        Δ s       Δ %
+  equilibrium                 2.10s      2.08s     -0.02     -1.0%
+  force_free_states         100.40s     98.10s     -2.30     -2.3%
+  total                     145.30s    142.80s     -2.50     -1.7%
 ```
 """
 
@@ -156,6 +163,28 @@ function checkout_ref(branch, commit)
     end
 end
 
+# Read the per-stage wall-clock seconds a run recorded under Info/Runtimes.
+# Returns an empty Dict for refs that predate the feature, so those still benchmark.
+function read_stage_runtimes(gpec_path)
+    stages = Dict{String,Float64}()
+    isfile(gpec_path) || return stages
+    h5open(gpec_path, "r") do h5
+        haskey(h5, "Info/Runtimes") || return
+        group = h5["Info/Runtimes"]
+        for stage in keys(group)
+            stages[stage] = read(group[stage])
+        end
+    end
+    return stages
+end
+
+# Average per-stage runtimes over the warm runs, keeping only stages present in every run.
+function average_stage_runtimes(stage_times)
+    isempty(stage_times) && return Dict{String,Float64}()
+    shared = intersect(map(keys, stage_times)...)
+    return Dict(stage => sum(s[stage] for s in stage_times) / length(stage_times) for stage in shared)
+end
+
 # Run the example benchmark.
 # Each run is a fresh Julia subprocess so that git branch switches take effect
 # (using statement only fires once per process; re-checkout without subprocess restart
@@ -176,19 +205,22 @@ function run_example_benchmark(example_path, num_runs)
         println("\n[1/$(num_runs+1)] First run (JIT compilation)...")
         run(`julia --project=$project_root $tmpscript $abs_example_path`)
 
-        # Warm runs for timing
+        # Warm runs for timing. Each run overwrites gpec.h5, so the per-stage
+        # Info/Runtimes record must be collected inside the loop, not after it.
+        gpec_path = joinpath(abs_example_path, "gpec.h5")
         runtimes = Float64[]
+        stage_times = Vector{Dict{String,Float64}}()
         for i in 1:num_runs
             println("\n[$((i+1))/$(num_runs+1)] Warm run $i...")
             t_start = time()
             run(`julia --project=$project_root $tmpscript $abs_example_path`)
             runtime = time() - t_start
             push!(runtimes, runtime)
+            push!(stage_times, read_stage_runtimes(gpec_path))
             println("  Runtime: $(round(runtime, digits=2)) s")
         end
 
         # Extract metrics from gpec.h5
-        gpec_path = joinpath(abs_example_path, "gpec.h5")
         if !isfile(gpec_path)
             error("gpec.h5 not found at $gpec_path")
         end
@@ -203,13 +235,31 @@ function run_example_benchmark(example_path, num_runs)
         return (
             eigenvalue=real(et[1]),
             steps=nsteps,
-            runtime=avg_runtime
+            runtime=avg_runtime,
+            stages=average_stage_runtimes(stage_times)
         )
     finally
         rm(tmpscript; force=true)
     end
 end
 
+# Per-stage runtime breakdown (Info/Runtimes), averaged over the warm runs of each branch.
+# Silently skipped when either side lacks the record, so comparisons against older refs work.
+function print_stage_comparison(r1, r2)
+    shared = intersect(keys(r1.metrics.stages), keys(r2.metrics.stages))
+    isempty(shared) && return nothing
+    ordered = sort(collect(shared); by=stage -> (stage == "total", stage))
+
+    println("\nPer-stage runtime (avg over warm runs):")
+    @printf("  %-22s %10s %10s %10s %9s\n", "Stage", "Branch 1", "Branch 2", "Δ s", "Δ %")
+    for stage in ordered
+        t1 = r1.metrics.stages[stage]
+        t2 = r2.metrics.stages[stage]
+        @printf("  %-22s %9.2fs %9.2fs %+9.2f %+8.1f%%\n", stage, t1, t2, t2 - t1, 100 * (t2 - t1) / t1)
+    end
+    return nothing
+end
+
 # Main benchmarking function
 function benchmark_branches(options)
     println("="^60)
@@ -293,6 +343,8 @@ function benchmark_branches(options)
     speedup_desc = time_delta < 0 ? "faster" : "slower"
     @printf("  Runtime Δ:      %+.2f s (%+.1f%% %s)\n", time_delta, abs(time_pct), speedup_desc)
 
+    print_stage_comparison(r1, r2)
+
     println("\n" * "="^60)
 
     # Write to output file if requested

From a496b731e1831f0adddb65e28876d13919de2896 Mon Sep 17 00:00:00 2001
From: logan-nc <6198372+logan-nc@users.noreply.github.com>
Date: Fri, 14 Aug 2026 17:36:26 -0400
Subject: [PATCH 3/9] DOCS - HANDOFF - Add machine-switch handoff document
 (delete before merge)

Temporary scaffolding for continuing this branch on another machine: what the
branch does, the merge resolution against upstream's inline-path convention,
verification already done, the wrong-baseline regression pitfall, and the
outstanding work. Carries its own removal instructions.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 HANDOFF_h5_runtime_records.md | 133 ++++++++++++++++++++++++++++++++++
 1 file changed, 133 insertions(+)
 create mode 100644 HANDOFF_h5_runtime_records.md

diff --git a/HANDOFF_h5_runtime_records.md b/HANDOFF_h5_runtime_records.md
new file mode 100644
index 000000000..c0e60777d
--- /dev/null
+++ b/HANDOFF_h5_runtime_records.md
@@ -0,0 +1,133 @@
+# HANDOFF — `feature/h5-runtime-records`
+
+**Temporary working document. Delete it before this PR merges — see "Remove this file" at the bottom.**
+
+Written 2026-08-14 when the working session moved to another machine. Everything described here is
+pushed to `origin/feature/h5-runtime-records`.
+
+## What this branch does
+
+`main_from_inputs` already timed every pipeline stage but only logged the durations. This branch
+persists them to `gpec.h5` under `Info/Runtimes/<stage>` (Float64 seconds, annotated with
+`long_name` + `units="s"`), and teaches the branch-benchmark tool to report per-stage deltas.
+
+Stages recorded, only when they actually run: `equilibrium`, `galerkin`, `force_free_states`,
+`slayer`, `perturbed_equilibrium`, `kinetic_forces`, plus `total`.
+
+**Runtimes are informational only** — machine- and load-dependent. They are not regression
+quantities and must never be added to the regression harness.
+
+Target: PR into `refactor/hdf5-metadata` (stack: #363 → #364 → this).
+
+## Commits
+
+| Commit | Contents |
+|---|---|
+| `df82c9e1` | `GPEC - NEW FEATURE` — `_write_runtimes!` + `_RUNTIME_LONG_NAMES`, six stage captures, both h5-bearing exit paths, doc schema row, two `runtests_fullruns.jl` assertions |
+| `e695e890` | `BENCHMARKS - IMPROVEMENT` — `read_stage_runtimes` / `average_stage_runtimes` / `print_stage_comparison` in `benchmarks/benchmark_git_branches.jl` |
+| `eb944f9b` | Merge of `origin/refactor/hdf5-metadata` (branch was ~60 commits behind) |
+
+### Design points worth knowing
+
+- The banner at both exit paths was rewritten to reuse `total_dt`, so the log line and the recorded
+  value cannot disagree.
+- `_write_runtimes!` guards on `isfile` (no-op when `write_outputs_to_HDF5=false`) and deletes an
+  existing `Info/Runtimes` group before writing, so reruns into an existing file stay idempotent.
+- Annotation goes through the existing `Utilities.HDF5Annotations.annotate!` — no new annotation
+  path was written. `test/runtests_h5_schema.jl` enforces this automatically: its metadata walker
+  fails if any `Info/Runtimes/*` dataset lacks `long_name`/`units`.
+- The equilibrium-only early exit is deliberately untouched — no h5 exists at that point.
+- In the benchmark tool the `Info/Runtimes` read happens **inside** the warm-run loop, because each
+  run overwrites `gpec.h5`. Refs predating the feature yield an empty dict and the per-stage table
+  is skipped silently, so old-vs-new comparisons still work.
+
+### Merge resolution note
+
+Upstream `refactor/hdf5-metadata` removed the shared `H5_*` path consts from
+`src/GeneralizedPerturbedEquilibrium.jl` in favour of inline literal paths with cross-reference
+comments. The merge conflicted there; it was resolved **upstream's way** — the `H5_RUNTIMES` const
+was dropped and `_write_runtimes!` now uses the `"Info/Runtimes"` literal directly. The second
+conflict was `galerkin_solve`'s changed signature (`wv=` replaced `vac_data=`); upstream's call was
+kept with the timing capture layered on top.
+
+## Verification already done
+
+All of the following ran **before** the upstream merge unless noted:
+
+| Check | Result |
+|---|---|
+| `runtests_h5_schema.jl` | 28/28 pass — confirms `Info/Runtimes/*` carries required metadata |
+| `runtests_fullruns.jl` | 19/19 pass (17 before; +2 new runtime assertions) |
+| Manual DIII-D run | all five recorded values match their `completed in` log lines exactly; `units`/`long_name` present; only stages that ran appear |
+| Benchmark dry-check | real read, missing-group fallback, missing-file fallback, shared-key averaging, `total`-last ordering, silent skip on empty/disjoint sides — all correct |
+| Regression harness, `diiid_n1`, `0ece9c4f` vs working tree | every physical quantity agrees to 1e-9–1e-16 (0.00%); see caveat below |
+| Package load after merge | clean |
+
+### Regression-harness caveat — read before rerunning
+
+The first harness run was done against `origin/refactor/hdf5-metadata` and reported **38 changed
+quantities**. That was a wrong-baseline artifact: at the time this branch's parent (`0ece9c4f`) was
+~60 commits behind that ref, so the report measured *upstream* physics changes (auto psi-grid Δ′
+convergence fix, on-demand solution derivatives, coil `psilim` fix) — not this work. Rerun against
+the branch's actual merge-base, not the remote branch tip.
+
+Against the correct baseline, 14 quantities were still flagged `** CHANGED **` but every one at
+**0.00%** (1e-9 to 1e-16 — floating-point noise; the harness flags any last-bit difference, and the
+profile "checksum" quantities flag on any bit change at all). Two residuals were being investigated
+when the session ended:
+
+- `ODE steps (total)` 1960 → 1968 (0.41%); `ODE steps (saved)` identical at 1327.
+- Three profile checksums (Mercier `D_I`, resistive interchange `D_R`, ballooning Δ′) differ.
+
+The diff cannot affect numerics — it computes `time()` differences and appends to the h5 *after* all
+computation — so the working hypothesis is run-to-run nondeterminism (threaded BLAS reassociation)
+and/or environment drift between the harness worktree and the local environment. A determinism check
+was running at handoff time: two `--force` runs of the *same* commit `0ece9c4f`, diffed against each
+other. **Its result was never seen.** Re-run it to close this out:
+
+```bash
+julia --project=regression-harness regression-harness/regress.jl --cases diiid_n1 --refs 0ece9c4f --force > a.log 2>&1
+julia --project=regression-harness regression-harness/regress.jl --cases diiid_n1 --refs 0ece9c4f --force > b.log 2>&1
+diff a.log b.log   # differences here ⇒ the case is nondeterministic ⇒ the residuals above are noise
+```
+
+## Outstanding work
+
+1. **Re-run the full test suite on the merged tree.** Only the package load and (in flight at
+   handoff) `runtests_h5_schema.jl` were checked after the merge. `runtests_fullruns.jl` (~23 min)
+   has not been re-run post-merge:
+   ```bash
+   julia --project=. test/runtests.jl runtests_h5_schema.jl runtests_fullruns.jl
+   ```
+2. **Close out the determinism check** above, and post the corrected regression table on the PR.
+3. **Optional:** a real two-branch benchmark run to see the per-stage table print end-to-end. Only a
+   dry-check against synthetic and real h5 files was done, by explicit choice — the full run costs
+   ~20–40 min of compute.
+4. **Remove this file** (below) and get human review.
+
+## Environment note
+
+If a `julia` invocation hits manifest errors on the new machine:
+`julia --project=. -e 'using Pkg; Pkg.resolve(); Pkg.instantiate()'`.
+**Never** remove a package from `Project.toml` — the developer works across several machines and
+environment drift is expected; fix the environment, not the manifest.
+
+## Remove this file
+
+This document is scaffolding for a machine switch and must not land in `refactor/hdf5-metadata`:
+
+```bash
+git rm HANDOFF_h5_runtime_records.md
+git commit -m "DOCS - CLEANUP - Remove the machine-switch handoff document"
+git push
+```
+
+Do this once the outstanding work above is finished and before the PR is approved for merge.
+
+---
+
+# ⚠️ **MERGE GATE** ⚠️
+
+# **NO PULL REQUEST IS EVER MERGED INTO `develop` WITHOUT A THIRD-PARTY HUMAN REVIEWER'S APPROVAL.**
+
+# **THIS IS NON-NEGOTIABLE. NO EXCEPTIONS.**

From e232b870de580134babe0014c1e19467f2fc5e70 Mon Sep 17 00:00:00 2001
From: logan-nc <6198372+logan-nc@users.noreply.github.com>
Date: Fri, 14 Aug 2026 17:49:41 -0400
Subject: [PATCH 4/9] DOCS - HANDOFF - Record post-merge verification and the
 closed regression question

Both test files pass on the merged tree, and the regression residuals are
resolved: the case is deterministic, and parent-vs-feature run through the same
worktree path shows 48 unchanged / 0 changed. Documents the commit-vs-local
environment trap that produced the earlier spurious diffs.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 HANDOFF_h5_runtime_records.md | 48 +++++++++++++++--------------------
 1 file changed, 20 insertions(+), 28 deletions(-)

diff --git a/HANDOFF_h5_runtime_records.md b/HANDOFF_h5_runtime_records.md
index c0e60777d..0a8422e50 100644
--- a/HANDOFF_h5_runtime_records.md
+++ b/HANDOFF_h5_runtime_records.md
@@ -56,11 +56,11 @@ All of the following ran **before** the upstream merge unless noted:
 
 | Check | Result |
 |---|---|
-| `runtests_h5_schema.jl` | 28/28 pass — confirms `Info/Runtimes/*` carries required metadata |
-| `runtests_fullruns.jl` | 19/19 pass (17 before; +2 new runtime assertions) |
+| `runtests_h5_schema.jl` | pass, **re-run after the merge** (14/14 + 6/6 group-name rule) |
+| `runtests_fullruns.jl` | 19/19 pass, **re-run after the merge** (17 before; +2 new runtime assertions) |
 | Manual DIII-D run | all five recorded values match their `completed in` log lines exactly; `units`/`long_name` present; only stages that ran appear |
 | Benchmark dry-check | real read, missing-group fallback, missing-file fallback, shared-key averaging, `total`-last ordering, silent skip on empty/disjoint sides — all correct |
-| Regression harness, `diiid_n1`, `0ece9c4f` vs working tree | every physical quantity agrees to 1e-9–1e-16 (0.00%); see caveat below |
+| Regression harness, `diiid_n1`, `0ece9c4f` vs `e695e890` | **48 unchanged, 0 changed** — zero movement |
 | Package load after merge | clean |
 
 ### Regression-harness caveat — read before rerunning
@@ -71,39 +71,31 @@ quantities**. That was a wrong-baseline artifact: at the time this branch's pare
 convergence fix, on-demand solution derivatives, coil `psilim` fix) — not this work. Rerun against
 the branch's actual merge-base, not the remote branch tip.
 
-Against the correct baseline, 14 quantities were still flagged `** CHANGED **` but every one at
-**0.00%** (1e-9 to 1e-16 — floating-point noise; the harness flags any last-bit difference, and the
-profile "checksum" quantities flag on any bit change at all). Two residuals were being investigated
-when the session ended:
+There is a second, subtler trap: **comparing a ref against `local` compares environments as well as
+code.** Non-`local` refs run in a freshly instantiated harness worktree; `local` runs the working
+tree with its own resolved environment. Run `0ece9c4f` vs `local` and 14 quantities come back
+flagged `** CHANGED **` — all at 0.00% (1e-9 to 1e-16), plus `ODE steps (total)` 1960 → 1968 and
+three profile checksums. That is environment drift, not code.
 
-- `ODE steps (total)` 1960 → 1968 (0.41%); `ODE steps (saved)` identical at 1327.
-- Three profile checksums (Mercier `D_I`, resistive interchange `D_R`, ballooning Δ′) differ.
+This was chased to ground and is now **closed**:
 
-The diff cannot affect numerics — it computes `time()` differences and appends to the h5 *after* all
-computation — so the working hypothesis is run-to-run nondeterminism (threaded BLAS reassociation)
-and/or environment drift between the harness worktree and the local environment. A determinism check
-was running at handoff time: two `--force` runs of the *same* commit `0ece9c4f`, diffed against each
-other. **Its result was never seen.** Re-run it to close this out:
+- *Is the case nondeterministic?* No. Two `--force` runs of the same commit `0ece9c4f` produced
+  byte-identical values for all 49 quantities; only the wall-clock line differed.
+- *Do these commits move any number?* No. Running the parent and the feature tip through the **same**
+  worktree path — `--refs 0ece9c4f,e695e890` — gives **48 unchanged, 0 changed**.
 
-```bash
-julia --project=regression-harness regression-harness/regress.jl --cases diiid_n1 --refs 0ece9c4f --force > a.log 2>&1
-julia --project=regression-harness regression-harness/regress.jl --cases diiid_n1 --refs 0ece9c4f --force > b.log 2>&1
-diff a.log b.log   # differences here ⇒ the case is nondeterministic ⇒ the residuals above are noise
-```
+So: compare worktree-to-worktree (two commit refs), not commit-vs-`local`, whenever the numbers
+need to be trusted at last-bit precision.
 
 ## Outstanding work
 
-1. **Re-run the full test suite on the merged tree.** Only the package load and (in flight at
-   handoff) `runtests_h5_schema.jl` were checked after the merge. `runtests_fullruns.jl` (~23 min)
-   has not been re-run post-merge:
-   ```bash
-   julia --project=. test/runtests.jl runtests_h5_schema.jl runtests_fullruns.jl
-   ```
-2. **Close out the determinism check** above, and post the corrected regression table on the PR.
-3. **Optional:** a real two-branch benchmark run to see the per-stage table print end-to-end. Only a
+1. **Optional:** a real two-branch benchmark run to see the per-stage table print end-to-end. Only a
    dry-check against synthetic and real h5 files was done, by explicit choice — the full run costs
    ~20–40 min of compute.
-4. **Remove this file** (below) and get human review.
+2. **Remove this file** (below) and get human review.
+
+Everything else is done: both test files pass on the merged tree and the regression comparison is
+clean.
 
 ## Environment note
 

From 31285e193d986dd6e36ac90fe549512870884a0c Mon Sep 17 00:00:00 2001
From: logan-nc <6198372+logan-nc@users.noreply.github.com>
Date: Fri, 14 Aug 2026 17:50:25 -0400
Subject: [PATCH 5/9] DOCS - HANDOFF - Point the harness caveat at the
 documented Manifest-pinning fix

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 HANDOFF_h5_runtime_records.md | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/HANDOFF_h5_runtime_records.md b/HANDOFF_h5_runtime_records.md
index 0a8422e50..e676532e5 100644
--- a/HANDOFF_h5_runtime_records.md
+++ b/HANDOFF_h5_runtime_records.md
@@ -87,6 +87,12 @@ This was chased to ground and is now **closed**:
 So: compare worktree-to-worktree (two commit refs), not commit-vs-`local`, whenever the numbers
 need to be trusted at last-bit precision.
 
+This is a known class of artifact — see `docs/development/regression-harness.md`, "Making source
+code the only variable": an unpinned `Manifest.toml` lets a worktree resolve different package
+versions, and the adaptive ODE step controller amplifies machine-epsilon library differences into
+apparent regressions. The harness merged in from upstream now pins the working tree's Manifest into
+every worktree; the misleading run above was made with the pre-merge harness, which predates that.
+
 ## Outstanding work
 
 1. **Optional:** a real two-branch benchmark run to see the per-stage table print end-to-end. Only a

From c170b9684705e6323520329bde609bf59897e265 Mon Sep 17 00:00:00 2001
From: logan-nc <nikolas.logan@columbia.edu>
Date: Sat, 15 Aug 2026 02:32:41 -0400
Subject: [PATCH 6/9] GPEC - IMPROVEMENT - Gate runtime records on this run
 writing the output file

The isfile guard alone let a run with write_outputs_to_HDF5=false stamp its
timings into a stale gpec.h5 left over from an earlier run. Gate both call
sites on the run having produced the file, keeping the SLAYER-created case.

Move the stage long_names into HDF5Schema.jl as RUNTIME_H5_ANNOTATIONS, next
to the other Info/ metadata; annotate! skips the stages that did not run.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dh1NCejnd3fYMmcRKoQRcG
---
 src/GeneralizedPerturbedEquilibrium.jl | 28 +++++++++++++-------------
 src/HDF5Schema.jl                      | 12 +++++++++++
 2 files changed, 26 insertions(+), 14 deletions(-)

diff --git a/src/GeneralizedPerturbedEquilibrium.jl b/src/GeneralizedPerturbedEquilibrium.jl
index 1359a7137..19306c7cd 100755
--- a/src/GeneralizedPerturbedEquilibrium.jl
+++ b/src/GeneralizedPerturbedEquilibrium.jl
@@ -556,7 +556,10 @@ function main_from_inputs(
         slayer_result = _run_slayer_stage(nothing)
         total_dt = time() - total_start
         push!(runtimes, "total" => total_dt)
-        _write_runtimes!(joinpath(intr.dir_path, ctrl.HDF5_filename), runtimes)
+        # A returned SLAYER result means that stage created or appended to the file itself.
+        if ctrl.write_outputs_to_HDF5 || slayer_result !== nothing
+            _write_runtimes!(joinpath(intr.dir_path, ctrl.HDF5_filename), runtimes)
+        end
         @info "\n$_BANNER\n  GPEC completed successfully in $(@sprintf("%.3f", total_dt)) s\n$_BANNER"
         return (ctrl=ctrl, equil=equil, intr=intr, ffit=ffit, odet=odet,
             free_energies=free_energies,
@@ -694,7 +697,10 @@ function main_from_inputs(
     # ----------------------------------------------------------------
     total_dt = time() - total_start
     push!(runtimes, "total" => total_dt)
-    _write_runtimes!(joinpath(intr.dir_path, ctrl.HDF5_filename), runtimes)
+    # A returned SLAYER result means that stage created or appended to the file itself.
+    if ctrl.write_outputs_to_HDF5 || slayer_result !== nothing
+        _write_runtimes!(joinpath(intr.dir_path, ctrl.HDF5_filename), runtimes)
+    end
     @info "\n$_BANNER\n  GPEC completed successfully in $(@sprintf("%.3f", total_dt)) s\n$_BANNER"
 
     # TODO: Do not allow perturbed equilibrium calculations if zero crossings are found
@@ -1030,22 +1036,16 @@ function _write_coil_snapshot!(h5_path::String, coil_sets::Vector{ForcingTerms.C
     return nothing
 end
 
-const _RUNTIME_LONG_NAMES = Dict(
-    "equilibrium" => "Wall-clock time of the Equilibrium construction stage",
-    "force_free_states" => "Wall-clock time of the ForceFreeStates stability stage",
-    "galerkin" => "Wall-clock time of the Galerkin outer-region solve",
-    "slayer" => "Wall-clock time of the SLAYER tearing-mode stage",
-    "perturbed_equilibrium" => "Wall-clock time of the PerturbedEquilibrium stage",
-    "kinetic_forces" => "Wall-clock time of the KineticForces (NTV) stage",
-    "total" => "Wall-clock time of the full GPEC run")
-
 """
     _write_runtimes!(h5_path::String, runtimes)
 
 Write the per-stage wall-clock seconds collected during a run into `Info/Runtimes/` of an
 existing gpec.h5 file. `runtimes` iterates `stage => seconds` pairs; only the stages that ran
-are recorded. These timings are informational only — machine- and load-dependent, never a
-regression quantity.
+are recorded. Metadata comes from `RUNTIME_H5_ANNOTATIONS`, which skips the absent stages.
+These timings are informational only — machine- and load-dependent, never a regression quantity.
+
+Call it only when this run produced the file; the `isfile` guard alone would also stamp a
+stale gpec.h5 left over from an earlier run.
 """
 function _write_runtimes!(h5_path::String, runtimes)
     isfile(h5_path) || return nothing
@@ -1054,7 +1054,7 @@ function _write_runtimes!(h5_path::String, runtimes)
         for (stage, dt) in runtimes
             out_h5["Info/Runtimes/$stage"] = dt
         end
-        Utilities.HDF5Annotations.annotate!(out_h5, ["Info/Runtimes/$stage" => (; long_name=_RUNTIME_LONG_NAMES[stage], units="s") for (stage, _) in runtimes])
+        Utilities.HDF5Annotations.annotate!(out_h5, RUNTIME_H5_ANNOTATIONS)
     end
     return nothing
 end
diff --git a/src/HDF5Schema.jl b/src/HDF5Schema.jl
index 047b452fb..d704741ef 100644
--- a/src/HDF5Schema.jl
+++ b/src/HDF5Schema.jl
@@ -185,6 +185,18 @@ const MAIN_H5_ANNOTATIONS = [
     "SurfaceGeometries/Wall/z" => (; long_name="Cartesian z of wall point cloud", units="m"),
 ]
 
+# Per-stage wall-clock records written by `_write_runtimes!` after every stage has run.
+# Stages that did not run leave their path absent and are skipped by `annotate!`.
+const RUNTIME_H5_ANNOTATIONS = [
+    "Info/Runtimes/equilibrium" => (; long_name="wall-clock time of the Equilibrium construction stage", units="s"),
+    "Info/Runtimes/galerkin" => (; long_name="wall-clock time of the Galerkin outer-region solve", units="s"),
+    "Info/Runtimes/force_free_states" => (; long_name="wall-clock time of the ForceFreeStates stability stage", units="s"),
+    "Info/Runtimes/slayer" => (; long_name="wall-clock time of the SLAYER tearing-mode stage", units="s"),
+    "Info/Runtimes/perturbed_equilibrium" => (; long_name="wall-clock time of the PerturbedEquilibrium stage", units="s"),
+    "Info/Runtimes/kinetic_forces" => (; long_name="wall-clock time of the KineticForces (NTV) stage", units="s"),
+    "Info/Runtimes/total" => (; long_name="wall-clock time of the full GPEC run", units="s"),
+]
+
 # Euler-Lagrange operator matrices: same wording per letter, Ideal/ and Kinetic/ variants.
 const _ELM_IDEAL_LETTERS = [
     ("A", "Euler-Lagrange primitive coefficient matrix A"),

From fe3127859e0b8bdc2d5d2c1e9c232a71ca222e31 Mon Sep 17 00:00:00 2001
From: logan-nc <nikolas.logan@columbia.edu>
Date: Sat, 15 Aug 2026 02:32:46 -0400
Subject: [PATCH 7/9] DOCS - CLEANUP - Remove the machine-switch handoff
 document

Temporary scaffolding for a machine switch; its contents are folded into the
PR description.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dh1NCejnd3fYMmcRKoQRcG
---
 HANDOFF_h5_runtime_records.md | 131 ----------------------------------
 1 file changed, 131 deletions(-)
 delete mode 100644 HANDOFF_h5_runtime_records.md

diff --git a/HANDOFF_h5_runtime_records.md b/HANDOFF_h5_runtime_records.md
deleted file mode 100644
index e676532e5..000000000
--- a/HANDOFF_h5_runtime_records.md
+++ /dev/null
@@ -1,131 +0,0 @@
-# HANDOFF — `feature/h5-runtime-records`
-
-**Temporary working document. Delete it before this PR merges — see "Remove this file" at the bottom.**
-
-Written 2026-08-14 when the working session moved to another machine. Everything described here is
-pushed to `origin/feature/h5-runtime-records`.
-
-## What this branch does
-
-`main_from_inputs` already timed every pipeline stage but only logged the durations. This branch
-persists them to `gpec.h5` under `Info/Runtimes/<stage>` (Float64 seconds, annotated with
-`long_name` + `units="s"`), and teaches the branch-benchmark tool to report per-stage deltas.
-
-Stages recorded, only when they actually run: `equilibrium`, `galerkin`, `force_free_states`,
-`slayer`, `perturbed_equilibrium`, `kinetic_forces`, plus `total`.
-
-**Runtimes are informational only** — machine- and load-dependent. They are not regression
-quantities and must never be added to the regression harness.
-
-Target: PR into `refactor/hdf5-metadata` (stack: #363 → #364 → this).
-
-## Commits
-
-| Commit | Contents |
-|---|---|
-| `df82c9e1` | `GPEC - NEW FEATURE` — `_write_runtimes!` + `_RUNTIME_LONG_NAMES`, six stage captures, both h5-bearing exit paths, doc schema row, two `runtests_fullruns.jl` assertions |
-| `e695e890` | `BENCHMARKS - IMPROVEMENT` — `read_stage_runtimes` / `average_stage_runtimes` / `print_stage_comparison` in `benchmarks/benchmark_git_branches.jl` |
-| `eb944f9b` | Merge of `origin/refactor/hdf5-metadata` (branch was ~60 commits behind) |
-
-### Design points worth knowing
-
-- The banner at both exit paths was rewritten to reuse `total_dt`, so the log line and the recorded
-  value cannot disagree.
-- `_write_runtimes!` guards on `isfile` (no-op when `write_outputs_to_HDF5=false`) and deletes an
-  existing `Info/Runtimes` group before writing, so reruns into an existing file stay idempotent.
-- Annotation goes through the existing `Utilities.HDF5Annotations.annotate!` — no new annotation
-  path was written. `test/runtests_h5_schema.jl` enforces this automatically: its metadata walker
-  fails if any `Info/Runtimes/*` dataset lacks `long_name`/`units`.
-- The equilibrium-only early exit is deliberately untouched — no h5 exists at that point.
-- In the benchmark tool the `Info/Runtimes` read happens **inside** the warm-run loop, because each
-  run overwrites `gpec.h5`. Refs predating the feature yield an empty dict and the per-stage table
-  is skipped silently, so old-vs-new comparisons still work.
-
-### Merge resolution note
-
-Upstream `refactor/hdf5-metadata` removed the shared `H5_*` path consts from
-`src/GeneralizedPerturbedEquilibrium.jl` in favour of inline literal paths with cross-reference
-comments. The merge conflicted there; it was resolved **upstream's way** — the `H5_RUNTIMES` const
-was dropped and `_write_runtimes!` now uses the `"Info/Runtimes"` literal directly. The second
-conflict was `galerkin_solve`'s changed signature (`wv=` replaced `vac_data=`); upstream's call was
-kept with the timing capture layered on top.
-
-## Verification already done
-
-All of the following ran **before** the upstream merge unless noted:
-
-| Check | Result |
-|---|---|
-| `runtests_h5_schema.jl` | pass, **re-run after the merge** (14/14 + 6/6 group-name rule) |
-| `runtests_fullruns.jl` | 19/19 pass, **re-run after the merge** (17 before; +2 new runtime assertions) |
-| Manual DIII-D run | all five recorded values match their `completed in` log lines exactly; `units`/`long_name` present; only stages that ran appear |
-| Benchmark dry-check | real read, missing-group fallback, missing-file fallback, shared-key averaging, `total`-last ordering, silent skip on empty/disjoint sides — all correct |
-| Regression harness, `diiid_n1`, `0ece9c4f` vs `e695e890` | **48 unchanged, 0 changed** — zero movement |
-| Package load after merge | clean |
-
-### Regression-harness caveat — read before rerunning
-
-The first harness run was done against `origin/refactor/hdf5-metadata` and reported **38 changed
-quantities**. That was a wrong-baseline artifact: at the time this branch's parent (`0ece9c4f`) was
-~60 commits behind that ref, so the report measured *upstream* physics changes (auto psi-grid Δ′
-convergence fix, on-demand solution derivatives, coil `psilim` fix) — not this work. Rerun against
-the branch's actual merge-base, not the remote branch tip.
-
-There is a second, subtler trap: **comparing a ref against `local` compares environments as well as
-code.** Non-`local` refs run in a freshly instantiated harness worktree; `local` runs the working
-tree with its own resolved environment. Run `0ece9c4f` vs `local` and 14 quantities come back
-flagged `** CHANGED **` — all at 0.00% (1e-9 to 1e-16), plus `ODE steps (total)` 1960 → 1968 and
-three profile checksums. That is environment drift, not code.
-
-This was chased to ground and is now **closed**:
-
-- *Is the case nondeterministic?* No. Two `--force` runs of the same commit `0ece9c4f` produced
-  byte-identical values for all 49 quantities; only the wall-clock line differed.
-- *Do these commits move any number?* No. Running the parent and the feature tip through the **same**
-  worktree path — `--refs 0ece9c4f,e695e890` — gives **48 unchanged, 0 changed**.
-
-So: compare worktree-to-worktree (two commit refs), not commit-vs-`local`, whenever the numbers
-need to be trusted at last-bit precision.
-
-This is a known class of artifact — see `docs/development/regression-harness.md`, "Making source
-code the only variable": an unpinned `Manifest.toml` lets a worktree resolve different package
-versions, and the adaptive ODE step controller amplifies machine-epsilon library differences into
-apparent regressions. The harness merged in from upstream now pins the working tree's Manifest into
-every worktree; the misleading run above was made with the pre-merge harness, which predates that.
-
-## Outstanding work
-
-1. **Optional:** a real two-branch benchmark run to see the per-stage table print end-to-end. Only a
-   dry-check against synthetic and real h5 files was done, by explicit choice — the full run costs
-   ~20–40 min of compute.
-2. **Remove this file** (below) and get human review.
-
-Everything else is done: both test files pass on the merged tree and the regression comparison is
-clean.
-
-## Environment note
-
-If a `julia` invocation hits manifest errors on the new machine:
-`julia --project=. -e 'using Pkg; Pkg.resolve(); Pkg.instantiate()'`.
-**Never** remove a package from `Project.toml` — the developer works across several machines and
-environment drift is expected; fix the environment, not the manifest.
-
-## Remove this file
-
-This document is scaffolding for a machine switch and must not land in `refactor/hdf5-metadata`:
-
-```bash
-git rm HANDOFF_h5_runtime_records.md
-git commit -m "DOCS - CLEANUP - Remove the machine-switch handoff document"
-git push
-```
-
-Do this once the outstanding work above is finished and before the PR is approved for merge.
-
----
-
-# ⚠️ **MERGE GATE** ⚠️
-
-# **NO PULL REQUEST IS EVER MERGED INTO `develop` WITHOUT A THIRD-PARTY HUMAN REVIEWER'S APPROVAL.**
-
-# **THIS IS NON-NEGOTIABLE. NO EXCEPTIONS.**

From 9bd0ef23cd9984278ad7817a8f168b1b0a58a0f3 Mon Sep 17 00:00:00 2001
From: logan-nc <nikolas.logan@columbia.edu>
Date: Sat, 15 Aug 2026 09:13:41 -0400
Subject: [PATCH 8/9] TEST - BUGFIX - Exempt Info/Runtimes from the replay
 bit-for-bit comparison

The replay guard compares every dataset in a source gpec.h5 against its rerun.
Wall-clock seconds cannot repeat, so the new records broke it the same way
Info/git_version would; skip the group alongside it.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dh1NCejnd3fYMmcRKoQRcG
---
 test/runtests_rerun_from_h5.jl | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/test/runtests_rerun_from_h5.jl b/test/runtests_rerun_from_h5.jl
index 2e364f2df..cb4420e6b 100644
--- a/test/runtests_rerun_from_h5.jl
+++ b/test/runtests_rerun_from_h5.jl
@@ -7,10 +7,11 @@ using TOML
 
 # Collect every leaf dataset path under an open HDF5 file, skipping the groups/paths that
 # legitimately differ between a source run and its replay (`Input/` is re-emitted with the
-# rerun's own filename/TOML blob; `Info/git_version` reflects the running commit).
+# rerun's own filename/TOML blob; `Info/git_version` reflects the running commit;
+# `Info/Runtimes/` records wall-clock seconds, which never repeat).
 function _rerun_leaf_paths(h5)
     skip_toplevel = Set(["Input"])
-    skip_paths = Set(["Info/git_version"])
+    skip_paths = Set(["Info/git_version", "Info/Runtimes"])
     paths = String[]
     function walk(node, prefix)
         for k in keys(node)

From 40e550e3c9ab86306ffad10f2cd9ec366468b81a Mon Sep 17 00:00:00 2001
From: logan-nc <nikolas.logan@columbia.edu>
Date: Sat, 15 Aug 2026 10:43:25 -0400
Subject: [PATCH 9/9] GPEC - IMPROVEMENT - Name the tearing runtime record and
 stamp the file the run wrote

Four corrections to the Info/Runtimes records after the schema overhaul landed:

- The stage token is `tearing`, not `slayer`. The `slayer/` group became `Tearing/`
  under the "name the physics, not the algorithm" rule, and every other token already
  mirrors its group. The solver is named in the long_name instead.
- `perturbed_equilibrium` is recorded only when the section is present. It was pushed
  unconditionally, so a stability-only run stamped the stage at ~0 s for work that
  never happened.
- `_write_runtimes!` now targets the file the run actually wrote. It hardcoded
  ctrl.HDF5_filename, but PerturbedEquilibrium, KineticForces and SLAYER each write
  their own filename, so timings were silently dropped whenever a stage other than
  ForceFreeStates produced the output. Each write site records its path in
  `written_h5` (last writer wins, matching where SLAYER appends) and both exit paths
  stamp that file, replacing the write_outputs_to_HDF5/slayer_result proxy with the
  fact it was approximating.
- The benchmark per-stage table prints an em dash instead of Inf/NaN when the
  branch-1 baseline is zero.

The galerkin and total long_names now state that the values nest rather than
partition the run, so a reader does not sum them, and workflow.md's Info/ row
mentions Runtimes/.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VGmFuAw5JdYrAyBSXCVssR
---
 benchmarks/benchmark_git_branches.jl   |  4 +++-
 docs/src/workflow.md                   |  2 +-
 src/GeneralizedPerturbedEquilibrium.jl | 25 +++++++++++++++++--------
 src/HDF5Schema.jl                      |  7 ++++---
 4 files changed, 25 insertions(+), 13 deletions(-)

diff --git a/benchmarks/benchmark_git_branches.jl b/benchmarks/benchmark_git_branches.jl
index 63367fb0e..3daf54d5d 100755
--- a/benchmarks/benchmark_git_branches.jl
+++ b/benchmarks/benchmark_git_branches.jl
@@ -255,7 +255,9 @@ function print_stage_comparison(r1, r2)
     for stage in ordered
         t1 = r1.metrics.stages[stage]
         t2 = r2.metrics.stages[stage]
-        @printf("  %-22s %9.2fs %9.2fs %+9.2f %+8.1f%%\n", stage, t1, t2, t2 - t1, 100 * (t2 - t1) / t1)
+        # A stage that ran in no measurable time has no meaningful percentage baseline.
+        pct = t1 == 0 ? "—" : @sprintf("%+.1f%%", 100 * (t2 - t1) / t1)
+        @printf("  %-22s %9.2fs %9.2fs %+9.2f %9s\n", stage, t1, t2, t2 - t1, pct)
     end
     return nothing
 end
diff --git a/docs/src/workflow.md b/docs/src/workflow.md
index 228459ed3..a39cec8d8 100644
--- a/docs/src/workflow.md
+++ b/docs/src/workflow.md
@@ -174,7 +174,7 @@ All results are written to a single HDF5 file (default: `gpec.h5`). The top-leve
 
 | Group | Contents |
 |---|---|
-| `Info/` | Run metadata: git version, mode-number ranges, ψ limit |
+| `Info/` | Run metadata: git version, mode-number ranges, ψ limit, `Runtimes/` (per-stage wall-clock seconds) |
 | `Input/` | Self-contained rerun snapshot: merged TOML blob, raw equilibrium/forcing/coil inputs |
 | `Equilibrium/` | Equilibrium scalars (`beta_N`, `q_axis`, `q_95`, …), 1-D profiles (`Profiles/`), 2-D geometry (`Geometry/`) |
 | `ForceFreeStates/` | Stability solve: `Solutions/{ForwardIntegration,GalerkinIntegration}`, `EulerLagrangeMatrices/`, `FreeBoundaryStability/`, `EdgeScan/` |
diff --git a/src/GeneralizedPerturbedEquilibrium.jl b/src/GeneralizedPerturbedEquilibrium.jl
index b3e431315..17c3ab413 100755
--- a/src/GeneralizedPerturbedEquilibrium.jl
+++ b/src/GeneralizedPerturbedEquilibrium.jl
@@ -174,6 +174,9 @@ function main_from_inputs(
     total_start = time()
     # Per-stage wall-clock seconds, written to Info/Runtimes at the end of the run.
     runtimes = Vector{Pair{String,Float64}}()
+    # Output file this run actually wrote, last writer wins (matching where the SLAYER stage
+    # appends). `nothing` means no stage produced a file, so there is nothing to stamp.
+    written_h5 = nothing
 
     # ----------------------------------------------------------------
     # Equilibrium
@@ -508,6 +511,7 @@ function main_from_inputs(
             locstab=locstab,
             ballooning_boundary=ballooning_boundary
         )
+        written_h5 = ctrl.HDF5_filename
         @info "Results written to $(ctrl.HDF5_filename)"
     end
 
@@ -532,7 +536,7 @@ function main_from_inputs(
             result = Runner.run_slayer(equil, intr, slayer_ctrl;
                 dir_path=intr.dir_path)
             slayer_dt = time() - slayer_start
-            push!(runtimes, "slayer" => slayer_dt)
+            push!(runtimes, "tearing" => slayer_dt)
             @info "SLAYER completed in $(@sprintf("%.3f", slayer_dt)) s"
             h5_filename = pe_file === nothing ? ctrl.HDF5_filename : pe_file
             h5_path = joinpath(intr.dir_path, h5_filename)
@@ -541,6 +545,7 @@ function main_from_inputs(
             HDF5.h5open(h5_path, isfile(h5_path) ? "r+" : "w") do f
                 Runner.write_slayer_hdf5!(f, result)
             end
+            written_h5 = h5_filename
             @info "SLAYER results written to $h5_filename"
             return result
         catch err
@@ -556,9 +561,8 @@ function main_from_inputs(
         slayer_result = _run_slayer_stage(nothing)
         total_dt = time() - total_start
         push!(runtimes, "total" => total_dt)
-        # A returned SLAYER result means that stage created or appended to the file itself.
-        if ctrl.write_outputs_to_HDF5 || slayer_result !== nothing
-            _write_runtimes!(joinpath(intr.dir_path, ctrl.HDF5_filename), runtimes)
+        if written_h5 !== nothing
+            _write_runtimes!(joinpath(intr.dir_path, written_h5), runtimes)
         end
         @info "\n$_BANNER\n  GPEC completed successfully in $(@sprintf("%.3f", total_dt)) s\n$_BANNER"
         return (ctrl=ctrl, equil=equil, intr=intr, ffit=ffit, odet=odet,
@@ -635,6 +639,7 @@ function main_from_inputs(
             PerturbedEquilibrium.write_outputs_to_HDF5(
                 pe_state, pe_intr, joinpath(intr.dir_path, output_file)
             )
+            written_h5 = output_file
             @info "Results written to $output_file"
         end
 
@@ -646,7 +651,11 @@ function main_from_inputs(
     end
 
     pe_dt = time() - pe_start
-    push!(runtimes, "perturbed_equilibrium" => pe_dt)
+    # Record only when the stage ran; an absent [PerturbedEquilibrium] section would otherwise
+    # stamp a ~0 s entry for work that never happened.
+    if "PerturbedEquilibrium" in keys(inputs)
+        push!(runtimes, "perturbed_equilibrium" => pe_dt)
+    end
     @info "Perturbed Equilibrium completed in $(@sprintf("%.3f", pe_dt)) s"
 
     # ----------------------------------------------------------------
@@ -672,6 +681,7 @@ function main_from_inputs(
                 h5open(joinpath(intr.dir_path, kf_ctrl.HDF5_filename), "cw") do h5file
                     KineticForces.write_to_hdf5!(h5file, kf_state; dVdpsi_spline=equil.profiles.dVdpsi_spline)
                 end
+                written_h5 = kf_ctrl.HDF5_filename
             end
         end
 
@@ -697,9 +707,8 @@ function main_from_inputs(
     # ----------------------------------------------------------------
     total_dt = time() - total_start
     push!(runtimes, "total" => total_dt)
-    # A returned SLAYER result means that stage created or appended to the file itself.
-    if ctrl.write_outputs_to_HDF5 || slayer_result !== nothing
-        _write_runtimes!(joinpath(intr.dir_path, ctrl.HDF5_filename), runtimes)
+    if written_h5 !== nothing
+        _write_runtimes!(joinpath(intr.dir_path, written_h5), runtimes)
     end
     @info "\n$_BANNER\n  GPEC completed successfully in $(@sprintf("%.3f", total_dt)) s\n$_BANNER"
 
diff --git a/src/HDF5Schema.jl b/src/HDF5Schema.jl
index ebb5f32e7..54b302cd7 100644
--- a/src/HDF5Schema.jl
+++ b/src/HDF5Schema.jl
@@ -245,12 +245,13 @@ const MAIN_H5_ANNOTATIONS = [
 # Stages that did not run leave their path absent and are skipped by `annotate!`.
 const RUNTIME_H5_ANNOTATIONS = [
     "Info/Runtimes/equilibrium" => (; long_name="wall-clock time of the Equilibrium construction stage", units="s"),
-    "Info/Runtimes/galerkin" => (; long_name="wall-clock time of the Galerkin outer-region solve", units="s"),
+    "Info/Runtimes/galerkin" => (; long_name="wall-clock time of the Galerkin outer-region solve (measured inside force_free_states, not additional to it)", units="s"),
     "Info/Runtimes/force_free_states" => (; long_name="wall-clock time of the ForceFreeStates stability stage", units="s"),
-    "Info/Runtimes/slayer" => (; long_name="wall-clock time of the SLAYER tearing-mode stage", units="s"),
+    "Info/Runtimes/tearing" => (; long_name="wall-clock time of the Tearing stage (SLAYER inner-layer solve)", units="s"),
     "Info/Runtimes/perturbed_equilibrium" => (; long_name="wall-clock time of the PerturbedEquilibrium stage", units="s"),
     "Info/Runtimes/kinetic_forces" => (; long_name="wall-clock time of the KineticForces (NTV) stage", units="s"),
-    "Info/Runtimes/total" => (; long_name="wall-clock time of the full GPEC run", units="s"),
+    "Info/Runtimes/total" =>
+        (; long_name="wall-clock time of the full GPEC run (the per-stage values nest rather than partition it, so they do not sum to this)", units="s"),
 ]
 
 # Euler-Lagrange operator matrices: same wording per letter, Ideal/ and Kinetic/ variants.