chore(search,error-lookup): execFileSync for rg + trim noisy code refs

[critesjosh]

Summary

Two related cleanups in the search / error-lookup stack, surfaced in a review of how MCP tool output reaches the calling LLM.

src/utils/search.ts — argv-based ripgrep, no shell layer

The old code interpolated filePattern and query into a shell-command string for execSync, which created two bugs:

• *.{nr,ts} was brace-expanded by /bin/sh and glob-expanded against the node-process cwd (not searchPath) before reaching ripgrep. Whether anything matched depended on which /bin/sh you happened to have. The same bug silently affected searchDocs with *.{md,mdx}.
• A query that started with - (e.g. searchCode("-g")) would be reparsed by ripgrep as a flag once the shell quoting layer was gone.

This switches to argv-based execFileSync. Patterns reach rg verbatim, the query goes through -e <query> followed by a -- end-of-flags marker, and escapeShell is removed (it was the only consumer of the old shell-string form).

src/utils/error-lookup.ts — trim noisy code references

When the static catalog has fewer than three matches, lookupError falls back to grepping *.ts across aztec-packages for related code. Up to five raw hits used to flow straight through to the formatter, regularly including .test.ts, fixtures, and minified ABI/bytecode lines that didn't help anyone debug the error.

The fix over-fetches (RAW_CODE_LIMIT = 20), filters via a new isUsefulCodeRef heuristic, and caps the emitted list to 2.

• Path-based exclusions are the primary signal: .test.ts, .spec.ts, .e2e.ts, plus segment-bounded __tests__ / tests / e2e / fixtures / mocks directories. Segment-bounded so real source dirs (attestations/, testdata/, latest/, contest/) survive untouched.
• Content-shape exclusion is deliberately narrow: drop only lines ≥400 chars that contain a 200+ char contiguous hex run. Real generated-looking code (long enum maps, regex literals, selector tables) is preserved — better to over-include than drop a legitimate "this is the file you want" pointer.

The filter is localized to lookupError rather than searchCode because the suppression-of-semantic-fallback gate in src/tools/error-lookup.ts reads codeMatches.length, and aztec_search_code callers may legitimately want test files anyway.

Test plan

• npm test — 306 tests pass (303 existing + 3 new test cases). New coverage:
  • search.test.ts: argv-shape assertions for single-extension and brace globs (the regression-fixing claim), -e <query> -- <path> ordering for flag-shaped queries, brace globs flowing unchanged through the manual-search fallback (globby/micromatch handles braces natively).
  • error-lookup.test.ts: isUsefulCodeRef and isMinifiedShape unit-by-unit, including negative controls for paths that contain test-like substrings (attestations/, testdata/, latest/, contest/), a long-but-not-hex-shaped generated source line that must survive, paths that begin with a test segment without a leading slash, and an integration test that pipes a synthetic rg stdout (5 tests + 1 minified + 2 real source) through lookupError and asserts both the two real refs survive and the over-fetch limit reaches rg as -m 40.
• npm run build (tsc) — clean.

Risk

• Behaviour change: anyone running with a /bin/sh that was accidentally leaving *.{md,mdx} alone (and so searchDocs was working) sees no change. Anyone running on a shell that was brace-expanding sees searchDocs start working. Net win.
• escapeShell removal: only used by the old call site; verified by grep before removal.
• isUsefulCodeRef false negatives: the heuristic prefers "keep" over "drop." Tests cover the obvious cases; the eval bucket of "user reports lookup_error gave them a useless test ref" should drop accordingly.

🤖 Generated with Claude Code

[github-actions]

🎉 This PR is included in version 1.21.2 🎉

The release is available on:

• npm package (@latest dist-tag)
• GitHub release

Your semantic-release bot 📦🚀