SIGSEGV in lsp_cross pass when all_defs exceeds ~1189 entries (scale-dependent, not file-specific)

[Codyzzz-zach]

Bug: SIGSEGV (exit 139) in lsp_cross pass when indexing TS/JS monorepos with 1189+ definitions

Summary

The lsp_cross pass crashes with SIGSEGV when indexing TypeScript/JavaScript monorepos where pxc_collect_all_defs produces a CBMLSPDef[] array exceeding ~1189 entries. The crash is scale-dependent, not file-specific — any subset of files that stays below the threshold indexes successfully.

Environment

• OS: macOS 15 (Darwin 24.0.0, arm64)
• Binary: v0.6.1 darwin-arm64 (prebuilt release)
• Project: Vue 3 + TypeScript monorepo (pnpm workspace)
  • 1832 files total, 4864 defs extracted
  • 899 TS/JS/TSX files requiring lsp_cross processing
  • Backend (Kotlin) indexes fine in full mode — Kotlin has no cross-file LSP, so pxc_has_cross_lsp() returns false

Reproduction

# Clone a Vue 3 + TS monorepo with 1800+ files
codebase-memory-mcp index --project my-frontend --repo /path/to/frontend --mode full
# → SIGSEGV at lsp_cross pass (exit code 139)

Scale Threshold Isolation

Systematic binary-search testing isolates the crash to the total defs count passed to cbm_run_ts_lsp_cross:

Scope	Files	Defs	Result
Minimal 2-file TS project	2	4	OK
apps/ alone	917	1160	OK
packages/ alone	471	1189	CRASH
packages/ui-kit alone	337	~600	CRASH
packages/ui-kit/shadcn-ui alone	251	25	OK
Any single packages/* sub-package	varies	<200	OK
Combined sub-packages (total defs >1189)	varies	1189+	CRASH

The crash triggers regardless of which specific files are included — it's the accumulated all_defs array size that matters.

Root Cause Analysis

The crash path is:

cbm_pipeline_pass_lsp_cross()           [pass_lsp_cross.c:400]
  → pxc_collect_all_defs()              [pass_lsp_cross.c:149]
      // Creates single shared CBMLSPDef[1189+] array
  → for each TS/JS file:
      → pxc_run_one_ts()                [pass_lsp_cross.c:381]
          → cbm_run_ts_lsp_cross()      [ts_lsp.c:4230]
              // Registers ALL defs into CBMTypeRegistry
              // → SIGSEGV

Key observations from source analysis (pass_lsp_cross.c):

1. pxc_collect_all_defs (line 149) collects ALL project definitions into a single CBMLSPDef[] array — every Class, Interface, Function, Method, Enum, Type, Protocol, Trait across all files.

2. cbm_pipeline_pass_lsp_cross (line 400) passes this entire all_defs array to every pxc_run_one_ts / pxc_run_one call. For 899 TS files, the same 1189+ def array is registered into 899 separate type registries.

3. pxc_run_one_ts (line 381) uses a per-file scratch arena, but the all_defs array and the type registry built from it are not isolated — memory corruption in one iteration can cascade.

4. The per-file scratch arena in pxc_run_one/pxc_run_one_ts was added to prevent O(N×project_size) memory growth (noted in comment at line 326-330), but it doesn't protect against corruption in the shared defs array or the type registry registration path.

Suggested Fix: Batch Processing

Split the TS/JS file loop in cbm_pipeline_pass_lsp_cross into batches. Each batch:

1. Collects only the definitions relevant to its subset of files (direct defs + imported defs)
2. Creates a fresh CBMLSPDef[] subset and scratch arena
3. Processes its file subset independently

This limits the all_defs array size per batch, avoiding the memory corruption threshold while preserving cross-file resolution within each batch.

Relationship to Other Issues

• Segmentation fault at start of lsp_cross pass on repositories with 200+ files #340 — Same lsp_cross SIGSEGV, filed 2 days ago with bisect pointing to May 8 LSP commits. This issue adds scale-threshold isolation data and C source-level root cause analysis.
• Dump-phase crash on large TS monorepo (v0.6.1, darwin-arm64) — pipeline completes through semantic_edges then terminates abnormally before gbuf.dump #317 — Dump-phase crash on large TS monorepo (138K files). May be a different manifestation of the same underlying memory safety issue.
• Crash on Large Java Project: Memory Corruption Causes Database Corruption #189 — Dump-phase crash with heap corruption signature (pointer authentication trap IB). The lsp_cross corruption could cascade into the dump phase.

Workaround

Index with mode: "moderate" (skips lsp_cross pass) or index sub-packages individually below the ~1189 def threshold.