v1.7.0

Fixed

• stdio audits now probe the stateless MCP 2026-07-28 lifecycle. Seven
  transport-independent JSON-RPC probes run on one short-lived, no-handshake
  process through the SDK's cross-platform stdio transport; the five
  HTTP-specific probes remain not applicable. A
  dual-era stdio server is therefore reported and scored as dual-era instead
  of being assigned an unavoidable protocol_version_latest failure and a
  misleading readiness result based only on unprobed checks.
• protocol_version_latest now judges observed lifecycle support. A legacy
  handshake capped at 2025-11-25 passes when gateway probes demonstrate
  2026-07-28 support, fails when those probes explicitly reject it, and skips
  when the modern gateway could not be observed.
• Readiness summaries disclose skipped checks. The CLI qualifies a readiness
  fraction with the number of checks assessed whenever any were skipped, and
  JSON reports expose readiness.assessed, readiness.skipped, and
  readiness.total.

Changed

• A partial audit no longer prints a bare N/N. An auth-gated server with a
  correct authorization posture scores 25/25, which read as a perfect result:
  the caveat printed above it does not travel when someone screenshots or pastes
  the number. The score line now carries the qualifier and the check count
  itself:

  Audit finished. PARTIAL score: 25/25 from 9 of 68 checks — not comparable to a full audit.
  

  The count is main-axis on both sides, matching the score it qualifies;
  readiness reports its own totals on its own line. Full audits are unchanged —
  they still print Audit finished. Final score: N/M.

Added

• --package: score an npm or PyPI package without running it. Half the MCP
  registry — 9,832 of 20,016 active servers in the 2026-08 crawl — ships as a
  package and speaks stdio, with no URL to point an auditor at.
  mcpscore --package npm:@scope/server, --package npm:name@1.2.3 and
  --package pypi:name==1.2.3 fetch the registry's published metadata and score
  how the server is packaged: whether it resolves, whether a pinned version
  exists, whether the release has been withdrawn (PyPI yank / npm deprecate),
  and whether it declares a source repository, a license and a description.

  The package is never downloaded and never executed — no tarball, no wheel,
  no install hook — so this is safe to run against anything, including packages
  you have not reviewed. It therefore says nothing about whether the server
  speaks MCP correctly; run the server with --stdio for that.

  No credential ever reaches a package registry. The fetch builds a bare
  request and disables client auth, so a caller's Authorization, cookies and
  custom headers cannot follow an audit to npm or PyPI — a guarantee that holds
  even when an authenticated client is passed in. The response is read with a
  hard size cap applied while streaming, so an oversized or endless document
  is refused rather than buffered.

  Six new rules in a packaging group. They are the only rules a package
  audit runs, and no server audit runs any of them: the two judge different
  targets, so their scores share no denominator and existing reports are
  unchanged. A dry run against 120 packages listed in the MCP registry measured
  pass rates of 100% (resolves), 99% (version resolves), 95% (not withdrawn),
  93% (license), 83% (repository) and 100% (description) — the pack is lenient
  by design at this stage, with the source-repository check the main
  discriminator. The JSON report gains a package block whose executed: false
  records what the score is and is not based on.