Replace vmem with mimalloc global allocator, add extraction-phase prescan, fix __init__.py QN collision

Memory management:
- Vendor mimalloc v2.1.9 as global allocator (MI_OVERRIDE=1 in prod)
- New mem.h/mem.c: RSS-based budget tracking via mi_process_info()
- Remove vmem.c/vmem.h (mmap-based budget tracking)
- Remove slab tier2 bump allocator (~300 LOC); >64B goes to mimalloc
- Slab tier1 pages from malloc (= mimalloc) instead of vmem
- Arena blocks from malloc instead of vmem
- Budget raised from 35% to 50% RAM (no more untracked C++ heap)

Extraction-phase prescan (eliminates disk re-reads):
- HTTP call sites: keyword check + URL extraction during extraction
- HTTP routes: decorator + source-based extraction during extraction
- Config file refs: regex scan during extraction
- httplinks: 41.8s → 13ms on Linux kernel (3,212x faster)
- configlink: 41.4s → 0.8s on Linux kernel (54x faster)
- Linux kernel fast-mode total: 2m38s → 1m18s

Bug fixes:
- __init__.py Module QN no longer collides with Folder QN
- index.ts same fix for JS/TS packages
- 13 regression tests for QN collision at FQN + extraction layers
- search_graph/search_code default limit raised from 10 to 500k
- Resolve all clang-tidy, cppcheck, and clang-format warnings

Repo cleanup:
- tree-sitter-form, tree-sitter-magma moved to tools/
- .gitignore: build/, node_modules/, graph-ui/dist/, TEST_PLAN.md