ci(release): check the release people can actually download

[donislawdev]

The verify command in both READMEs did not work, and 0.5.0 shipped it.

$ gh attestation verify BeanNetworkTester-v0.5.0-windows-x64.zip -R donislawdev/BeanNetworkTester
Error: HTTP 404: Not Found (.../attestations/sha256:f6b8bbfe...?predicate_type=https%3A%2F%2Fslsa.dev%2Fprovenance%2Fv1)

gh defaults to asking for build provenance. The archive a user downloads deliberately has
none: a person signs it on their own machine, and attest-release.yml refuses to claim a
runner built it. What ships beside the release is the SBOM attestation over the signed bytes,
so the command has to say so. Measured, not assumed - adding
--predicate-type https://spdx.dev/Document/v2.3 makes both the online and the offline form
exit 0 against the real published release.

This is the second time that command has been wrong, and both times a person found it after
the fact, because nothing in the suite runs gh. So the text is only half the fix.

What changed

The documentation

• All three occurrences across README.md and README.pl.md now name the predicate type, and
  the prose no longer promises a proof the file does not carry.

A workflow that checks what people download (verify-release.yml, on release: published)

• Runs the README's own commands, verbatim, against the published assets - not equivalents, so
  the documentation cannot drift away from something that works.
• Then: all four assets present, sha256sum -c, Get-AuthenticodeSignature with status Valid
  and a timestamp, the signing certificate's SHA-256 against the pin read out of
  beantester/legal.py, and, for a full release, that /releases/latest is this tag - the
  site's download button points there.
• contents: read and nothing else. Re-runnable by workflow_dispatch for any published tag.

Two manual rules become gates

• release.yml refuses to build unless CI succeeded on exactly the tagged commit. It builds and
  publishes but does not test, so this was the last large step of the recipe kept by memory. It
  asks for the CI workflow by name rather than for a clean sweep, because one check in this
  repository fails for a licensing reason of its own and would otherwise block every release.
• sign_release.py acts on the certificate's expiry instead of printing it: a warning inside 90
  days, a refusal after it, both naming CODESIGN_SHA256 because a renewal is a different
  certificate and the pin moves with it.

And a seventh step for phase B

• sign_release.py now waits for the attestation workflow and confirms the draft is complete:
  four assets, the checksums downloaded back from the release naming the digest that was
  actually signed, still a draft. It used to end at "dispatched, go look" - and a draft missing
  one file looks almost exactly like a finished one.

Verification

• tests/test_version_and_release.py - 28 passed (25 before; three new guards).
• The widened guard on the documented commands found the third occurrence in each README within
  a minute of being written, which is the shape of guard worth having.
• Both documented commands run against the published v0.5.0: exit 0 with the predicate type,
  HTTP 404 without it.
• The PowerShell step's logic run over the real published archive: Valid, timestamped by
  Certum, certificate SHA-256 equal to the pin.
• All three release workflows parse as YAML; the new one triggers on release: published,
  holds contents: read, and runs on Windows because Authenticode exists nowhere else.
• ruff and mypy clean on the changed Python.

The full suite runs on CI rather than locally.