Segmentation fault at start of lsp_cross pass on repositories with 200+ files

[jjoos]

Description

After the May 8-9 LSP rework (commits adding py_lsp, php_lsp, TypeScript hybrid resolver, and the lsp_cross pipeline pass), codebase-memory-mcp segfaults immediately at the start of the lsp_cross pass when indexing medium-to-large repositories.

Reproduction

The crash happens consistently on any repository with ~200+ files. Observed on:

• project 1: 268 files, 13,671 extracted nodes — crashes every time
• project 2: 1,290 files — crashes after semantic_edges pass (heap corruption from earlier)

Logs

level=info msg=parallel.extract.done nodes=13671 errors=0
level=info msg=pass.timing pass=parallel_extract elapsed_ms=130
level=info msg=parallel.registry.start files=268
level=info msg=parallel.registry.done entries=889 defines=13671 imports=206
level=info msg=pass.timing pass=registry_build elapsed_ms=14
level=info msg=pass.start pass=lsp_cross files=268
Segmentation fault

The crash occurs within ~20ms of lsp_cross starting — before any per-file work is logged — suggesting an initialization bug (NULL dereference, uninitialized pointer, or allocation failure).

Bisect

• 10a9b83 (May 8, before feat(pipeline): consume LSP-resolved calls in pass_calls): also crashes, but on a different repo and pass — suggests the bug may be in the underlying LSP resolver infrastructure landed just before
• Commits before 2e042da (May 8, first Python LSP commit): likely safe but not yet confirmed
• 06f7e0ee (April 30, pre-LSP): fully working — all repos index cleanly

Environment

• Architecture: linux/arm64 (Docker, OrbStack on Apple M-series)
• Build: compiled from source with gcc on Ubuntu 24.04
• Confirmed reproducible across 3+ clean builds with --no-cache

Related

PR #339 fixes a separate but related issue (remaining fixed-size stacks in extract_channels.c) — that fix alone does not resolve this crash.