elf-mcp: MCP server for embedded toolchain flash/RAM analysis #62

Progress update (AI: big-pickle / opencode):

tools/elf-mcp/ is functionally complete:

• toolchain.py: arm-none-eabi binutils resolver (ArmGNUToolchain install tree, /opt/homebrew/bin, $PATH) + runner protocol.
• parsers.py: pure-text parsers for objdump -h, readelf -S, nm -S, linker-script MEMORY{}, GNU ld .map (memory config, per-object blame, discarded input sections, function-sections format, load-address rows, debug/meta-section exclusion).
• tools.py: 9 tools (list_elf_files, binary_size, section_breakdown, memory_map, largest_symbols, per_object_map, verify_no_heap, disassemble, addr2line) + CLI entrypoint.
• mcp_server.py: JSON-RPC 2.0 stdio loop matching pylink-square-mcp protocol; wired into opencode.json.
• tests/: 31 pytest cases, pass on host without an ARM toolchain (fake runner + fixture text).
• Validated on the real M7 demo ELF: binary_size flash=0xe0f4 / ram=0x37b20 matches arm-none-eabi-size exactly; per-object map totals within ~2% of flash; memory_map shows FLASH 2.68% fill, DTCM 100%; verify_no_heap passes clean.

Remaining: run scripts/build-all-presets.sh + host/cross cmake workflows, then PR against develop.

Implemented in PR #63: 9 tools (binary_size, memory_map, per_object_map, largest_symbols, section_breakdown, verify_no_heap, disassemble, addr2line, list_elf_files). All host/cross workflows green; 31 pytest cases pass; validated against the M7 demo ELF (flash 0xe0f4 / RAM 0x37b20 == arm-none-eabi-size).