mirror of
https://github.com/DeusData/codebase-memory-mcp.git
synced 2026-11-11 09:25:35 +08:00
Andrew Hundt
e02c6fb34f
Make duplicate graph-buffer upserts and worker merges choose richer source spans for definition-like nodes, so full and containment rebuilds no longer depend on worker order for declaration-vs-definition locations. Pass C/C++ macro extraction mode through extraction contexts instead of relying on process-global pipeline state, preventing MCP runs with mixed modes from leaking macro-node policy across pipelines. Recompute configlink-owned CONFIGURES edges deterministically by clearing only key_symbol/dependency_import strategy edges, preserving direct CONFIGURES links, and sizing candidate arrays from discovered graph counts instead of fixed caps. Reuse cbm_path_base for manifest basename handling and make it recognize both POSIX and Windows separators. Validation: git diff --check; ASan/UBSan test-runner rebuild /private/tmp/cbm-pan4-20260702T112332Z-build-test-runner-after-basename.log; str_util 65/65 and configlink 9/9 /private/tmp/cbm-pan4-20260702T112531Z-affected-strutil-configlink.log; product build /private/tmp/cbm-pan4-20260702T112545Z-cbm-build-after-configlink.log; MCP self-dogfood one-source canonical equality /private/tmp/cbm-pan4-20260702T112614Z-one-source-mcp-after-configlink.json; graph_buffer 63/63 and extraction 256/256 /private/tmp/cbm-pan4-20260702T112822Z-affected-graphbuffer-extraction.log; incremental 161/161 /private/tmp/cbm-pan4-20260702T112851Z-incremental-suite.log. Signed-off-by: Andrew Hundt <ATHundt@gmail.com>