NEXT.md

Engine follow-up

The four previous priorities are implemented in the current working tree:

• Compiler-derived reference questions complement declaration questions. The benchmark reports precision/recall, empty/unavailable answers, latency and response volume separately for static codeindex, codeindex with LSP and Serena.
• MCP concise: true extends to references, callers, file symbol overviews and symbol indexes without changing default answers or result membership.
• callers <name> --lsp and MCP callers with lsp: true append incoming-call evidence, capability status and reasons for degradation. Multilanguage declarations are routed to their own configured servers.
• MCP initialization advertises available/active profiles. The existing playground lists the Node-only semantic/LSP variants with an explanation.

The engine remains a deterministic static indexer. Type-aware results depend on an explicitly configured server and its project coverage. Reference occurrences are not call edges; agreement differences are evidence to investigate, not proof that one tier is correct. The controlled callers/impact/search tests are local product checks and do not author the answer key for competitor comparisons.

See docs/engine-validation-2026-09-07.md for executed checks, measured reference coverage, fixes, Docker validation and remaining limitations. New measurements must retain corpus revision/tool versions and report unavailable measurements.