v3.10.0

Selective MCP tool filtering, WASM build-speed regression fix, and Haskell parity restoration. A new mcp.disabledTools config field lets you remove specific MCP tools from the schema entirely — useful for smaller-context models that don't need all 33 tools competing for the initial prompt budget. The 3.9.6 expansion of AST_TYPE_MAPS to 23 languages had a side effect of making WASM full builds re-parse every WASM-parseable file in the corpus; the per-file needsFn filter now scopes the re-parse correctly, dropping the 744-file dogfooding build from 14.0s back to 7.8s (matching the 3.9.5 baseline). A second parity fix gates astTypeMap lookups with Object.hasOwn so Haskell's constructor node type no longer walks the prototype chain to Object.prototype.constructor — restoring the Haskell resolver from 0%/0% precision/recall in 3.9.6 to 100%/33% (matching the 3.9.4 baseline). The release benchmark workflow has also been restructured: regression guards now run inside publish.yml before npm publishes, instead of after the docs PR lands, so a regression can no longer ship to npm and then fire on unrelated dev commits.

Features

• mcp: add mcp.disabledTools config to remove specific tools from the MCP schema — drops disabled tools entirely from the schema (not just rejected at runtime) so smaller-context models save initial-prompt tokens; tool names are normalized for matching (#1035)

Bug Fixes

• parity: gate astTypeMap lookup with Object.hasOwn — Haskell constructor nodes (Left, Right, Just, …) no longer fall through to Object.prototype.constructor, which was dropping the non-cloneable Object() function into astNodes.kind and crashing the worker boundary with function Object() { [native code] } could not be cloned; Haskell resolver returns to v3.9.4 baseline (precision=1.0, recall=0.333) (#1041)

Performance

• wasm: scope ensureWasmTrees re-parse to files that actually need it — wasm-worker-entry.ts now serializes empty astNodes arrays (empty ≠ undefined) and ensureWasmTrees accepts an optional needsFn filter so only files genuinely lacking data are re-parsed; WASM full build on the 744-file dogfooding corpus drops from 14.0s back to 7.8s, restoring the 3.9.5 baseline (#1038)

CI

• release: gate npm publish on benchmark regressions — moves the regression guard into a pre-publish-benchmark job in publish.yml so a regression fails the publish workflow before npm sees the new version, instead of firing on unrelated dev commits after the post-publish benchmark PR lands (#1040)
• bench: rename auto-generated benchmark branch prefix from benchmark/ to chore/ — aligns with the local guard-git.sh allow-list so post-publish benchmark PRs no longer require hook bypass when pushed from a Claude Code session (#1044)