fix: the swept dt values must be distinct on disk

[sotashimozono]

Description

production.toml swept the natural halving sequence, and the natural halving sequence is the
wrong answer. A swept float reaches its directory name through %.2f.

Type of Change

• 🐛 Bug Fix (bug)

Proposed Changes

dt = [1e-2, 5e-3, 2.5e-3, 1.25e-3] renders as 0.01, 0.01, 0.00, 0.00. Twenty distinct points
claim ten directories; half the sweep silently overwrites the other half while the ledger reports
every point as done. ParamIO.diagnose:

config
smoke.toml	ok=true	1 point
debug.toml	ok=false	4 points → 2 directories
production.toml	ok=false	20 points → 10 directories

Only smoke.toml was clean — and smoke.toml is the one CI runs. That is why a template could
ship a sweep that cannot be run at size.

The replacement values are not free either. MyModule.decay_error takes
n = round(Int, t_end/dt), so a dt that does not divide t_end integrates to a different final
time and contributes an error unrelated to the step size. [0.08, 0.04, 0.02, 0.01] is distinct
on disk and still wrong:

dt=0.08   n=12   n*dt=0.960   err=5.79e-04   err/dt=0.0072
dt=0.04   n=25   n*dt=1.000   err=2.03e-02   err/dt=0.5085
dt=0.02   n=50   n*dt=1.000   err=1.01e-02   err/dt=0.5042
dt=0.01   n=100  n*dt=1.000   err=5.02e-03   err/dt=0.5021

At dt=0.08 the truncation error cancels against the end-time error, so the coarsest point
lands below the finest and the convergence plot reads backwards.

[0.1, 0.05, 0.02, 0.01] satisfies both constraints — distinct under %.2f, exact divisors of
1.0 — and the rendered name equals the value rather than a rounding of it. debug.toml becomes
[0.05, 0.01] for the same reasons.

Usage or Results

# From a fresh copy of the template
cd projects/ExampleSweep
julia --project=. scripts/compute.jl configs/debug.toml
julia --project=. scripts/collect.jl configs/debug.toml

  a        dt         rel_error
  0.5      0.01       1.253e-03
  0.5      0.05       6.336e-03
  1        0.01       5.021e-03
  1        0.05       2.553e-02

  4 points

on-disk key directories:
  sysa0.50_numdt0.01   sysa0.50_numdt0.05   sysa1.00_numdt0.01   sysa1.00_numdt0.05

Four points, four directories, four distinct rows, and err/dt flat across them. All three
configs now report ok=true. test/runtests.jl and report/report.jl both still pass.

The upstream half

DataVault #39 and #41 fix the other side: Vault now honours [datavault] float_format = "auto"
— which was previously read by ParamIO.diagnose and ignored by the writer, so setting it turned
the diagnostic green while the data kept colliding — and warns at construction when a grid
collides at all. With that, a sweep can use whatever values the physics wants.

This template pins the registered DataVault 0.7, so it cannot adopt auto until there is a
release. Changing the values is the fix that works today, and it is the right example regardless:
a template's configs should not need a knob to be correct.

check list

• ./setup.sh を通したか — configs only, no renamed identifiers touched
• projects/ExampleSweep が instantiate できるか — compute / collect / report all re-run
• 新しいファイルを足したなら、setup.sh の書き換え対象に入っているか — no new files
• 施設固有の記述・資格情報を含めていないか

🤖 Generated with Claude Code

[github-actions]

📚 Docs preview: https://qatlashub.github.io/templateHPC.jl/previews/PR8/

(updates on each push to this PR)