Cache toolchain probes and keep GC out of parsing

[jserv]

Loading a large Kconfig tree spends most of its time somewhere other than parsing. scripts/Kconfig.include defines cc-option, as-instr, ld-option and friends in terms of $(shell,...), so a full load forks a hundred shells that each fork a compiler, and the cyclic garbage collector walks a heap that only grows while the parse builds it. This caches the probe results, always for the lifetime of a Kconfig instance and, when KCONFIG_SHELL_CACHE names a file, across runs as well, and suppresses collection for the duration of the parse.

The cache cannot simply key on the command. $(python,...) runs in this process with os in scope, so a Kconfig file can move os.environ or the working directory partway through a parse, and a result measured before that must never be reused after it. Comparing the raw environment mapping costs about a microsecond, cheap enough to do before every lookup, while the working directory needs a syscall and so is re-read only when in-process Python may have run. Only results measured in the context the cache file is stamped with are written to it.

Collection is suppressed by zeroing the gen0 threshold rather than by calling gc.disable(), so gc.isenabled() keeps reporting what the caller and the Kconfig files believe: a file running $(python,import gc; gc.disable()) still has that survive the parse. The full collection that reclaims a previously dropped tree runs only when the previous parse allocated enough to have left real work behind, since it walks the host application's whole heap and Kconfiglib is usually embedded in a build script.

The remaining commits are the tooling half. menuconfig's jump-to search lived inside a two-hundred-line curses function and could not be tested; extracting it and running the same cases through both tools turned up a real divergence, where guiconfig left stale matches behind after a regex failed to compile. The row renderers and the MENUCONFIG_STYLE parser had no tests either. CI now measures coverage, and .ci/check-format.sh asks git for its file list instead of walking the tree with find, which had it formatting vendored checkouts and __pycache__.

Verified on an Apple M1, macOS 15.6, CPython 3.14.7. python -m pytest tests/ is 308 passed, 8 skipped; ruff check . and .ci/check-format.sh are clean; the suite also passes with tkinter absent (262 passed, 53 skipped) and with a broken Tk install. Parsing a synthetic 3001-file tree goes from 2.15 s to 1.50 s, best of seven interleaved runs in fresh processes. A 200-probe tree goes from 0.77 s cold to 0.002 s with the cache warm. Peak RSS across six repeated loads stays at about 370 MiB against 445 MiB before. .config output is byte-identical with the cache cold, filling, and warm.

Two limitations are deliberate rather than overlooked. Variables that describe the invocation instead of the toolchain, MAKEFLAGS and MAKELEVEL among them, are left out of the cache fingerprint, because including them means a cache filled by make menuconfig never survives a recursive make; the cost is that a command reading one of them directly gets the answer from whichever invocation filled the cache. And os.putenv() writes the process environment without updating os.environ, so it is invisible to the context check. Both are documented in the module docstring.

---

Summary by cubic

Caches toolchain probe results and keeps the cyclic garbage collector out of parsing, cutting load time on a 3000-file tree from 2.15s to 1.50s. Also fixes UI search bugs and adds tests, coverage, and a benchmark.

Probe cache and GC

• Probe results for $(cc-option) and friends are memoized per Kconfig instance; setting KCONFIG_SHELL_CACHE persists them across runs, with a cache key covering environment and working directory so only results measured in the stamped context are reused.
• Probe functions are identified by identity rather than membership, so custom KCONFIG_FUNCTIONS callables with raising __hash__ or __eq__ no longer crash the parse.
• GC is suppressed during parsing via a zeroed gen0 threshold; a full collection runs only after a large parse, sized by measured allocation, to reclaim dropped trees.

UI and tooling

• menuconfig's jump-to search is extracted into a testable function; guiconfig now clears stale matches after a bad regex.
• Added tests for row renderers, style parser, search across both tools, and the probe cache.
• CI measures coverage with the pass floor set from runner measurements, since local runs on Python 3.14 score differently; the format script uses git's file listing and filters to files on disk.
• New benchmark script times load and UI phases, outputting plain text or JSON; it skips the GUI phase when tkinter is absent.

^{Written for commit a91b888. Summary will update on new commits.}