fix: repair the dead websiteUrl and guard the registry entry

[a-essawy]

The registry publish step is continue-on-error by design, since npm is the source of truth. The side effect is that anything wrong with server.json degrades to a warning in a log nobody reads.

What that already cost us

websiteUrl pointed at https://rendobar.com/docs/mcp/. That page moved to /docs/mcp-server when rendobar/docs#59 consolidated the MCP docs, and the old path now answers 405:

https://rendobar.com/docs/mcp/          HEAD=308  GET=405
https://rendobar.com/docs/mcp-server    HEAD=200  GET=200

That dead URL is live in published registry version 1.7.1 right now, and it is what aggregators show. package.json homepage had the same stale value. Both fixed here.

Nothing in CI would have caught it, and nothing would have caught the next one either.

The guard

scripts/check-registry-entry.mjs, its own CI job on every PR. It checks the things that break a publish after npm has already gone out:

• server.json name vs package.json mcpName — a mismatch is a hard publish rejection
• the three version fields release-please bumps via extra-files, against each other
• packages[0].identifier vs the npm package name
• description against the registry's 100-char cap
• a remotes entry exists, so the hosted server cannot silently vanish from the listing again
• every URL in the entry resolves

The URL check matters most for icons: they point at the brand CDN, synced from the monorepo, so this repo would never notice them going dead.

No dependencies. Node has fetch, and the failures that actually happen here are cross-file drift and dead URLs, not exotic schema violations — those get caught below.

Post-publish verification

release.yml now polls the registry after publishing until the released version is present and marked isLatest, failing the job if it never is. That routes into the existing rollback-on-failure job, which opens an issue. npm has already published at that point, so the fix is always forward — re-run once the cause is addressed.

Test plan

Guard, negative-tested by mutating a scratch copy. Baseline passes; every branch fails:

case	result
baseline	passes
version drift	exit 1
name != mcpName	exit 1
description 101 chars	exit 1
remotes removed	exit 1
non-https icon	exit 1

Verification step, extracted from the workflow and run against the live registry:

case	result
1.7.1 (live, latest)	exit 0
1.7.0 (present, superseded)	exit 1
9.9.9 (absent)	exit 1

Also: all three workflow files parse as YAML, tsc --noEmit clean, eslint clean, 72/72 tests pass.

Two bugs found and fixed while testing the guard itself — a missing fetch timeout that hung the run past 120s, and process.exit() firing with a pending AbortSignal timer, which trips a libuv assertion on Windows and returns 127 instead of 1.