feat(extract): add R language extractor (regex, no tree-sitter grammar exists)

[mkazbekov]

Summary

• .r/.R is in CODE_EXTENSIONS but has no AST extractor, so R files silently contribute nothing and trip the R (.r/.R) listed in CODE_EXTENSIONS but has no AST extractor — silently drops all R files with zero warning #1689 warning on every run. tree-sitter-r isn't published on PyPI, so this follows apex.py's precedent: a regex extractor over a comment/string-masked copy of the source, brace-matched so nested closures scope correctly.
• extract_r (graphify/extractors/r.py) emits function definitions (including backtick-quoted infix operators like `%||%`), calls, library()/require()/pkg:: imports, and source() edges — including unrolling the common for (mod in files) source(paste0(mod, ".R")) bootstrap loop into individual per-file edges, which a literal-string source() regex alone would miss.
• R has one flat namespace once every file is source()'d into .GlobalEnv, so a single-file extractor can't tell a cross-file call from a typo. extract_r resolves same-file calls and stashes a bare-name hint on the rest; _resolve_r_bare_calls (graphify/extractors/resolution.py), wired into extract() the same way _resolve_cross_file_imports is for Python, does the corpus-wide lookup afterward and leaves ambiguous (>1 candidate) names dangling rather than guessing — never inventing an edge.
• _symbol_safe_name moves to base.py (shared, not R-specific) because an all-symbol identifier normalizes to the empty string in make_id, which silently collapsed a %op% operator node onto its own file's node id and produced a self-loop contains edge. Fixed at the shared layer since any future backtick-identifier language hits the same bug.

Validation

Ran against a real 15-file, ~150-function R analysis pipeline (not a synthetic fixture):

• 97% symbol recall (138/142) against a prior LLM-extracted graph of the same codebase, 0 false-positive nodes
• 0 self-loops, 0 dangling sources edges — all 13 edges of a dynamic source() loop recovered correctly
• Every remaining dangling calls edge audited by hand: genuine external package calls (Biobase::, ggplot2::, base-R strrep/setNames/etc.), zero false negatives

Two existing tests pinned .r/.R as "no extractor exists" as their fixture (test_extract_warns_on_code_files_with_no_ast_extractor, test_extract_progress_final_line_uses_consistent_denominator) — closing the gap makes both fail by design. Repointed both at a synthetic .zzznolang extension so they keep testing the general no-extractor warning path without pinning it to a specific language, and added test_extract_r_extracts_functions_calls_and_sources to lock in the new behavior (functions, cross-file call resolution, imports, sources).

Full suite, this branch vs. clean HEAD: identical 54 pre-existing failures on both (tree-sitter-hcl unavailable, Windows watch/repo-root path handling, a shallow-clone git-log fixture — none R-related). Net +1 passing test, 0 regressions.

Test plan

• pytest tests/test_extract.py -k "no_ast_extractor or extract_r_extracts or no_warning_when_all_code or progress_final_line" — all pass
• Full suite diffed against clean HEAD — same failure set, +1 net pass
• Manual validation against a real-world R corpus (see above)
• Maintainer: confirm .r dispatch placement/_DISPATCH ordering convention is acceptable (placed next to .jl, the closest thematic neighbor — bespoke single-extension scripting extractor, no interop family)