fix(c02): don't score framework package-registry update-checks as egress overreach (litmus-v15)

[RubenSousaDinis]

Fixes #110.

Problem

The C-02 egress baseline (DEFAULT_EGRESS_BASELINE) was empty, so any outbound connection during probe 2.2 counted as the server's overreach. But many MCP frameworks/package tools phone a public registry at startup with zero author intent — chiefly FastMCP, whose stdio run() banner calls check_for_newer_version() → GET https://pypi.org/pypi/fastmcp/json (default on, cold cache in the sandbox). Every FastMCP-based server was therefore graded D for egress it never wrote.

Verified end-to-end: re-grading a FastMCP server with --server-env FASTMCP_CHECK_FOR_UPDATES=off flips it D→A (C-02 passes), proving pypi.org was the sole finding and it was framework noise. See #110 for the full trace.

Change

• Seed DEFAULT_EGRESS_BASELINE with public package-registry infrastructure: pypi.org, files.pythonhosted.org, registry.npmjs.org (new exported PACKAGE_REGISTRY_BASELINE).
• Deliberately narrow: registry APIs + download CDNs only. The cloud instance-metadata endpoint is NOT allowlisted (real SSRF/credential target) — it stays flagged.
• Bump METHODOLOGY_VERSION → litmus-v15 (framework-only egress moves D→A is a pass/fail-semantics change, consistent with how prior C-02/C-04 FP fixes bumped).

Tests

• egress-allowlist.test.ts: baseline now allows the three registry hosts; still rejects the metadata endpoint and un-named registry subdomains (exact hosts only).
• c02-egress.test.ts: a server whose only egress is pypi.org (incl. the in-sandbox :8443) under the default baseline → C-02 pass; egress to 169.254.169.254 under the default baseline → still fail.
• litmus.test.ts: methodology version assertion updated to litmus-v15.
• Full non-live probes suite green (434 passed); core, probes, litmus typecheck clean.

Notes for review

• Scope check: the empty baseline was previously intentional (comment noted it avoids "silently turning failing grades into passes"). This change makes that a deliberate, versioned decision rather than a silent one — hence the v15 bump. Happy to split the version bump out if you'd rather land it separately.
• The public /methodology page (separate repo) will need a v15 sync as a follow-up; not in this repo's tree.
• Not marking ready-to-merge — leaving the methodology-version call and the allowlist membership to your review.