Trim derivable content from CLAUDE.md

[rdotsch]

Two blocks of CLAUDE.md were content a fresh session can reconstruct on its own, so every session paid for them and got nothing back. 21,396 → 17,827 chars, ~892 est. tokens saved per session.

Every gotcha, rationale and convention is kept. Two items that happened to live inside the cut region were relocated rather than deleted (below).

Cut A — the devtools command block (18 lines)

roxygen2::roxygenise()
devtools::load_all()
devtools::test()
devtools::check()
devtools::install()

These are the standard invocations for any R package, inferable from DESCRIPTION + R/ + tests/testthat/ alone. Replaced by one line saying the usual workflow applies unchanged — which also fixes a lead-in sentence (Typical workflows…:) that would otherwise dangle.

Kept: the generateStimuli2IFC() parallel-worker gotcha (workers library(rcicr) themselves, so the package must be installed, not just load_all()-loaded, or they fail with "there is no package called 'rcicr'") and the ChangeLog convention.

Cut B — the architecture walkthrough (29 lines)

The two per-function subsections and the ASCII data-flow diagram. This is not merely derivable from the code — it duplicates README.md, which covers the same pipeline, the same .Rdata contract and the same flow in "How it works" (line 64) and "Anatomy of the .Rdata file" (line 113). Two copies of the same explanation drift, and the copy a reader hits first is then the wrong one.

Replaced by a pointer to those two README sections. Verified both headings exist, and that nothing anywhere in the repo cross-referenced the removed subsections or the diagram.

Kept: the section intro (the .RData-is-the-only-link contract) and all of "Notes on conventions in this codebase".

Two items relocated, not deleted

Both sat inside cut B's line range rather than in "Notes on conventions", so a straight deletion would have lost them. They are now bullets under "Notes on conventions":

• The generateCI.R scaling pointer — that none/constant/matched/independent is a user-facing decision documented at length in the roxygen header, to be read before changing scaling logic.
• The ref_lookup trap — that it looks like a cache for avoiding resimulation and is not one, having been empty since 2018 because its rows were measured under the pre-erratum infoVal formula and were correctly commented out when 01e547e adopted the Euclidean norm.

That second one is the reason this PR saves ~892 est. tokens rather than the ~1,156 first estimated. Worth it: a "looks like a cache, isn't" note cannot be recovered by reading the code, which is the whole test for what belongs in this file.

Not proposed

Migrating "Releases and versioning" (38 lines) to a lazily-loaded skill. Its content is correctness traps that fire exactly once, at release time — "never put a version number in a NEWS.md section heading" cost a real R CMD check NOTE on 2026-07-28 — and a skill that fails to trigger pays that cost twice. At 44% of the ~40,000-char warning threshold there is no pressure justifying the risk.

Checked

• Every remaining section is gotchas, rationale or repo etiquette. Spot-verified after the rebase that the parallel-worker failure, ref_lookup, pre_0.3.0/generator_version, R/zzz.R globals, the parallel-not-future convention, the scaling pointer and the ChangeLog convention all survive.
• The content Write down the code conventions, and log tidyverse style as a v2 item #156 added to this file — the five required-status-check contexts and the 0xC0000409 Windows-flake note — was re-tested against the same derivability bar and kept. The flake note is a specific observed failure with its resolution; the context list is enumerable via gh api but records intent, and its surrounding gotcha (the classic branches endpoint reports empty and misleads you) is the non-derivable part. One reservation for later: a list stamped "as of 2026-07-29" is a drift risk, and if it ever goes stale the fix is to keep the endpoint pointer and drop the enumeration rather than re-date it.
• CLAUDE.md is .Rbuildignored, so nothing here can reach the built tarball, and the file is on the reproducibility gate's inert allowlist — compare should post green without doing work.

---

Generated by Claude Code