Windows/MSYS2-gcc15: self-built main hangs at scale; self-built v0.11.0 crashes (mimalloc corrupted free list) — released binary is fine #2606

Summary

Two build-related reliability problems observed on this Windows box (32 GB RAM, MSYS2 UCRT gcc 15.2.0, Windows 11 x64), neither reproducible with the released 0.11.0 binary on the same machine:

1. The current main (4d9592b, after the v0.11.0 tag) hangs during indexing at scale. On a 1,820-file mixed workspace, v0.11.0 completes a fast index in 11 seconds; main (self-built, same flags below) was abandoned after 30 minutes with the last worker log ending mid-run. On a 370-file Rust-only tree the hang reproduces quickly:

   level=info msg=pass.start pass=structure files=307
   level=info msg=pass.done pass=structure nodes=369 edges=367
   <no further progress; worker eventually marked outcome=hang>
   

   and the supervisor quarantines files with phase=hang_reason=quarantined on unrelated file types (logs/kb_test/narrative_with_kb_close_ES1.md, web/openalgo-chart/src/components/Topbar/Topbar.module.css, a .py inside crates/), restarting repeatedly (118 worker_failed records over ~90 minutes).

2. v0.11.0 self-built with MSYS2 UCRT gcc 15 crashes at large scale. The released 0.11.0 binary (CI toolchain) indexes the same 1,820-file workspace in 11 s; a self-built v0.11.0 (tag 8972ea6) crashes deterministically (~5 min in, during parallel.resolve, right after cross_lsp.rust_registry scale=types=6559_funcs=20787):

   mimalloc: error: thread 0x71B0AE5000: corrupted free list entry of size 16b at 0x01DD0E220010
   exit code -1073740791 (0xC0000409 STATUS_STACK_BUFFER_OVERRUN)
   

   Reproduced on 3 projects (1,440–33k nodes); small projects (12–71 s) index fine. The corruption is heap-level (a 16-byte free-list entry), so this smells like a latent UB/gcclib difference — quite possibly in the vendored allocator override interplay (-Wl,--wrap=malloc + mimalloc) rather than in the graph code, and quite possibly also the cause of (1) if HEAD grows similar pressure.

Build command used (only deviation from release: -Wno-error instead of -Werror, MSYS2 UCRT gcc 15.2):

make -j8 -f Makefile.cbm cbm BUILD_DIR=build/c-mingw \
  CFLAGS_COMMON="-std=c11 -D_DEFAULT_SOURCE -D_GNU_SOURCE -Wall -Wno-error -Wdate-time \
  -Wno-unused-parameter -Wno-sign-compare -Wno-format-truncation -Wno-unused-result"

Also: internal/cbm/doc_pdf.c:1337 uses %zu in cbm_arena_sprintf, which fails -Werror=format under MSYS2 (unknown conversion type character 'z') — either the format attribute disagrees with the actual printf backend on MinGW, or the arena printf needs %zu support. A -Wno-format build passes.

Happy to test candidate patches (I can build both debug and release variants here).