field-test

# Field test Read `CLAUDE.md`, source and the full diff. Build the target list from every changed behavior, including complete pre-existing changes. Test the local build, not a globally installed or remote published version. Run `bun run devcheck`, `bun run rebuild`, and `bun run test:package`. Tests use known synthetic fixtures; do not execute the user's local server config or start unrelated services. Verify applicable boundaries: - All four capability classes, full optional descriptor metadata, server identity/instructions and negotiated protocol version. - Legacy default without discovery, auto fallback, pinned modern success/failure, and both stdio/Streamable HTTP. - Later pages, empty pages with cursors, empty-string cursors, repeated cursors, limits and later-page errors. Counts must disclose incompleteness. - Unadvertised, successful empty, malformed and failed lists. One server/list failure must not discard successful neighbors. - Config shape validation, disabled entries, transport conflicts and static headers. - Credential sentinels in env/args/headers/URLs and nested errors; none may reach diagnostics, callbacks or reports through configuration/error handling. - Rendering flags independently and together, supplied examples, zero connection duration, delimiter escaping and code-fence balance. - Startup failure, timeout, report-write failure and transport cleanup; no stranded processes or timers. - Compiled Node CLI help/version/options and programmatic import, with actual package contents inspected. For new bugs, observe a failing regression before fixing. For existing untested behavior, pin a passing characterization before editing. Use additional live targets only when the user explicitly requests them. Report commands and concrete outcomes, not inferred success.