Review WS5: Split god-files (mcp.rs, main.rs, build.rs), add lib.rs, error-handling consistency

[willem445]

From the July 2026 codebase review (see docs/plan-codebase-review-recommendations.md). Priority: MEDIUM (structural; best done after the workspace + perf issues to avoid churn).

Tasks

• Split src/sempkg/src/mcp.rs (3,291 lines) into mcp/{jsonrpc,schema,format,query,context}.rs — JSON-RPC transport loop, tool schemas, formatters, the UnifiedHit query pipeline (RRF/dedup/KWIC/merge), and McpContext + simple tool methods respectively.
• Split src/sempkg/src/main.rs (2,535 lines) into commands/{add,sync,query,...}.rs. Dedupe first: BuildOptions construction is copy-pasted between add_from_github (main.rs:1662-1716) and add_from_local (main.rs:2298-2348) with an identical resolve_dirs closure repeated 6×; extract resolve_dirs() + a single build_and_install_bundle().
• Split src/sembundle/src/build.rs (1,498 lines) into build/{codegraph,lance_docs,code_index,chunk,tools}.rs. While there: remove the dead _cwd param and never-used passthrough=false branch in invoke (build.rs:1383,1407-1417) — and note CodeGraph subprocess failures currently capture no stderr; fix that. Factor the near-identical chunk-emission tails of the two symbol extractors (build.rs:703-734 vs :900-929).
• Add src/lib.rs to sempkg (currently bin-only) re-exporting modules, making main.rs a thin wrapper — unlocks tests/ integration tests for the currently untested store.rs, tool dispatch, and install/sync orchestration.
• Error-handling consistency: delete 7 never-constructed SempkgError variants (error.rs: ManifestNotFound, PackageNotFound, InvalidBundle, NotIndexed, RerankerModelNotFound, Db, Reranker); pick one convention per layer (typed only where callers match on variants, else anyhow + .context()). In sembundle, add PackError::Lance/Db variants — InvalidField is currently abused for LanceDB/DB failures (build.rs:400,410,419,591,608,625,1070,1105,1121,1131), producing misleading "invalid field" errors. Consider consolidating the per-subcommand error enums (they're erased to Box<dyn Error> in main.rs anyway).

Verification

Pure refactor — full Rust test suite green, MCP functional suite green, no CLI output changes (CLAUDE.md requires output stability).