[krlmlr]

This PR was prepared with Claude Code (an LLM), as the template asks to disclose.

Stance. The glue (everything in src/ outside src/vendor/, including the Stimulus-generated src/rinterface.c and src/cpp11.cpp) must compile without warnings under -Wall -Wextra -pedantic -Wno-unused-parameter -Wno-cast-function-type. Warnings in the vendored igraph C library are counted but never fail anything. They get fixed upstream and arrive through the series. A warning is attributed to the file it points at, so a glue file is not blamed for what an igraph header says. This follows the model duckdb-r adopted in duckdb/duckdb-r#2620 and refined in duckdb/duckdb-r#2652, minus the vendored-scope enforcement, because rigraph has no patch stack for warning fixes. The reasons for each flag are in the new handbook/build/warnings/README.md.

What changes.

• scripts/warnings.sh compiles just the ten glue objects (about 6 s with ccache) with the gate's flags and fails on a glue warning. scripts/warnings.sh vendored compiles the rest and only reports. The flags reach that one compile through a makefile of its own, never src/Makevars (CRAN rejects -Wno-* there) or ~/.R/Makevars, which R CMD check would also read. duckdb-r turned its check red that way before jarl #2652.
• rcc: a new step in custom/after-install runs it in the smoke test only.
• each: a new warnings gate in scripts/rcc-one.sh, between clean and check, with a 600 s stage budget. It applies only to commits that carry this change. A vendored commit that brings a new warning inside src/vendor/ stays green. One that makes the glue warn gets an R-side repair like any other breakage.
• Stimulus (tools/py-stimulus): an optional input passed by value is now zero-initialized, and an output without a C counterpart no longer gets an unused SEXP. src/rinterface.c is regenerated, with a 13-line diff. This subrepo fix belongs upstream in igraph/stimulus as well.
• Handbook: the new leaf handbook/build/warnings/, linked from build/ and from the root. The docs-consistency check is clean.

Warnings today (GCC 13.3, full build).

• Glue: 7, all in the generated src/rinterface.c: five -Wmaybe-uninitialized on optional arguments and two -Wunused-variable in R_igraph_is_bipartite_coloring. This PR fixes them in the generator. The three ISO C function/object pointer conversions in src/rinterface_extra.c are left out of the report, the same exemption R CMD check makes. igraph keeps an R closure behind igraph_function_pointer_t, so the glue cannot avoid them.
• Vendored: 40 C/C++ warnings (15 -Wsign-compare, 13 -Wmaybe-uninitialized, 5 -Wunused-value from the flex lexers, plus a few others) and 144 Fortran warnings in ARPACK.
• None of these is one of R CMD check's "significant warnings", so CRAN's -Wall -pedantic run is clean on that count today. The existing check step already fails on such a WARNING wherever it occurs, vendored code included.

Verification. Locally, the glue gate fails on main's rinterface.c (it lists the seven warnings) and passes on this branch. EACH_GATES=warnings scripts/rcc-one.sh passes, as does scripts/each-shard-test.sh (20/20). CI's GCC 15 might find warnings that 13 does not, and this PR's own CI run is the first check of that.

• By submitting this pull request, I assign the copyright of my contribution to The igraph development team.

---

Generated by Claude Code