Add a test_run tool: run tests and return structured failures

[A-x6]

Test runs through the bash tool hand the agent a wall of raw output; it wastes tokens and misreads failures. Add a test_run tool ('test_run' item in the Agents with better hands roadmap group):

• Detects the project's test command (package.json test script with the right package manager via lockfile, cargo test, go test ./..., pytest), or takes an explicit command.
• Runs it and parses failures into structured entries (name, file, message) for pytest, go test, cargo test, bun test, jest/vitest/mocha glyphs, and TAP, plus pass/fail counts.
• Returns compact output: status, exit code, counts, failure list, and the raw output tail only on failure.
• Same bash permission action so existing allow/deny rules apply.