Brand UI, Excel spreadsheet link, worked example and a test suite

[saudzahirr]

Six commits. The app had no icon, no tests, no examples, and no way to touch a
spreadsheet; it now has all four.

Application icon

setWindowIcon appeared nowhere in the repo, so the app inherited the Python
interpreter's icon. The crystal-ball logo now ships inside the package at eight
resolutions plus a Windows .ico, installed as both the application and window
icon, with an AppUserModelID so the task bar shows it. Verified from a clean
wheel install.

Interface

vatic/theme.py replaces the hand-written XP-era style sheet with a token
system and a contrast auditor. Body ink is a neutral near-black at 15.6:1;
the four logo hues are reserved for the primary action, selection, panel
washes and the charts. Every audited ink-on-surface pairing clears WCAG AA,
and a test fails the build if that stops being true.

Layout: a branded header carrying the logo and the active analysis, and with it
the Run Simulation button, which previously existed only as F5 and a menu
entry. Group boxes that rendered as empty frames got titles, the tables got
add/remove actions, and the input sidebar scrolls -- the window's minimum
height drops from 1197px to 523px, so it fits a 1366x768 laptop.

Formulas are edited in a dialog rather than in a cell editor a few characters
wide. JetBrains Mono is bundled under the SIL OFL so the interface renders the
same on machines that do not have it.

Charts serve plotly.js from a bundled copy through Plotly.react against a
persistent page, so they no longer wait on a CDN round-trip (6.3s measured
here) and work with no network at all. All 13 chart types and the PDF/PPTX
exports moved onto the brand palette.

Maths

compute_statistics now also reports variance, skewness and kurtosis, which
the reference implementation reports per variable and this one did not. The
conventions match it deliberately: population variance, the standardised third
moment with no sample-size correction, and Pearson kurtosis where a normal
distribution sits at 3.0.

Runs are reproducible. Nothing seeded numpy's global state, which mcerp samples
through, so two runs of the same model returned different Cpk, different PPM
and a different tornado ordering. There is now a Seed control, defaulting to
"random" so behaviour is unchanged unless a seed is chosen.

Excel link

vatic began as a spreadsheet application -- Excel was the model, Python only
drove the sampling -- and that capability was missing here entirely.

vatic/sheetmodel.py is the pure data model and imports on every platform.
vatic/excel/ is the Windows half: typed errors with a worksheet-error
decoder, a session owning the COM connection, and a runner.

The original wrote one cell and recalculated once per trial: a cross-process
call per cell per iteration, about eighteen minutes for ten thousand trials.
The runner writes the whole sample matrix to a hidden sheet, points each
assumption cell at its column through INDEX, and builds a one-variable data
table over a trial index, so Excel resolves every trial in a single
recalculation. Measured against live Excel 16.0: 1.4s for ten thousand
trials, 0.45s to re-run.

Three details were settled by probing Excel rather than by assumption:

• Workbook.Calculate does not exist; the Application and Worksheet forms do.
• An unfilled cell reports Color 16777215 with Pattern xlNone, so restoring
  the colour alone leaves a white fill behind.
• A data table's input cell must live on the table's own sheet, or Excel
  rejects it outright.

Restoration runs in a finally block, so a crash or a cancel still puts the
workbook back, and it restores the original formula rather than its
computed value -- the original snapshotted the value and wrote that back,
silently destroying a formula in any tagged input cell. CoUninitialize is
deliberately not called on close: callers still hold Range proxies, and tearing
the apartment down beneath them turns their collection into
RPC_E_DISCONNECTED, which crashes the interpreter instead of raising.

A Spreadsheet menu connects a workbook and tags the cell selected in Excel
as an assumption or a forecast; tagged cells are tinted in the sheet and their
original fill is restored. Help > Using vatic with Excel documents the
whole flow, what a run changes, what it restores, and what each failure means.

Examples and tests

Neither existed before.

examples/seal_tolerance.py runs the Double-D seal gland tolerance stack-up
from the original project through this app's API. build_seal_workbook.py
regenerates the same model as a spreadsheet with named ranges. The upstream
repository carries no licence, so its workbook is not redistributed; the
generator rebuilds an equivalent one from the same published dimensions and
formulas, which also keeps it reproducible from source control.

113 tests, checked against closed-form values and independent derivations
rather than against the implementation. Ten drive a private hidden Excel
instance and skip where Excel is absent, so they never touch an open session.
A new Tests workflow runs Linux, Windows and macOS on 3.12 and 3.13.

Notes for review

• The Excel run happens on the GUI thread, so the window is unresponsive for
  the ~1.4s a run takes. A COM apartment belongs to its creating thread and a
  correct threaded version needs message pumping; at 1.4s that trade seemed
  right, but it is worth doing properly before anyone points this at a heavy
  workbook.
• The reference implementation's capability metrics could never have run: it
  calls ss.norm(...) without importing scipy.stats, relying on a
  from mcerp import * leak that modern mcerp closed with __all__. Its other
  defects -- the normal density used where the CDF is required, a Cpm
  exponent written with bitwise XOR, and an ignored zshift leaving Zst equal
  to Zlt -- are not reproduced here, and a test pins down each one.

🤖 Generated with Claude Code