Add a CLAUDE.md with dev commands and architecture notes

Claude Code sessions in this repo currently start with no repo-specific context, so each one re-derives the same non-obvious architecture by reading across R/: the attribute-based data contract, the long possible_data layout, and the fact that the foo.surv / foo.mgcv functions are not S3 methods.

Add a root CLAUDE.md covering:

• The dev command loop (devtools::load_all/document/test/check, single-test invocation) and what CI actually runs.
• Architecture that spans several files: the import/simulate -> fit_EM() -> plot pipeline; the attribute contract (scale, source, mic_class, model) that downstream functions read instead of taking arguments; the long possible_data layout (one row per observation x component, backtick-named probability columns, P(C=c|y,t) as the weight); the mu/pi model split; the full vs reduced approach; and the cross-validation chain.
• The dispatch gotcha: there is no UseMethod anywhere in the package, so fit_mu_model.mgcv and friends only look like S3 methods. Adding a backend means editing every branching wrapper.
• Recurring failure modes, notably roxygen examples (the source of most of the recent R CMD check fixes) and that U+2264 in MIC strings is data syntax rather than a typo.

Also add ^CLAUDE\.md$ and ^\.claude$ to .Rbuildignore, so neither becomes a non-standard top-level file in the build.

Start by reading the R/ directory, tracing the import/simulate -> fit_EM() -> plot pipeline, and checking the devtools command loop against what CI runs. Write the root CLAUDE.md with the listed architecture, dispatch and failure-mode notes, then update .Rbuildignore with both patterns and verify the documented commands and build-ignore changes.