Release 0.81.0 — the door an author could not find

[JimmyWesley]

Release 0.81.0 — the door an author could not find

One spec version, v0.81, and the package moves 0.80.0 → 0.81.0. It comes
from the shape of the release before it: 0.80.0 shipped the extension
mechanism and left the person it was built for with no way in.
docs/extending.md still cited spec v0.20, so the one document naming
extensions contradicted what had just been published; nothing anywhere
taught how to write one; and the console's install box asked for a path.
A path is on the host — through a browser that is the container's
filesystem, so an operator holding an extension they had just written had
no route to a remote Station short of publishing it to git first. The
local-file door served only somebody with a shell there, who would have
used the CLI.

What changed

An uploaded archive is a source (L.2). A fifth door into the same
resolver, not a fifth resolver: bytes arrive as {name, b64}, the shape
ingest has carried since J.8, so a browser that can already send a document
sends an extension with no new mechanism. It is unverified by
construction — there is no forge and no index identity behind an upload,
so nothing could have been verified, and the tier says that rather than
implying a check happened. It is refused before it is written when it
exceeds MONKEYLLM_STATION_EXT_UPLOAD_MAX_MB (25): this door is for an
extension's own code, dependencies are resolved at install, and a
vendored-wheel bundle belongs to the local-file route.

Every seam declares its handler's contract (L.3, L.9 rule 1). What a
curation handler receives, what an events handler receives and what each
returns were prose in the specification and nowhere in the code, so the
conformance kit could answer it works about an extension whose handler
took the wrong parameters — found on the first ingest, at whatever hour that
ran. The kit now checks the signature at install, and it does so by reading
the author's source without importing it: a heavy handler's module
imports the very dependency L.5 exists to keep out of this process, so a
check that imported would break the rule protecting the deployment in order
to catch a typo, and would run third-party code at kit time, which L.2 rule
5 is careful to say the kit does not do.

The rule hinges on defaults, not **kwargs. The first cut said kwargs
excused a misspelled parameter and was wrong in the exact case it was
written for: given on_event(event, forrest, **rest), **rest absorbs the
forest the host passes and forrest is still unfilled. The specification
paragraph was corrected to match the code, not the reverse.

The authoring surface (L.16). Prose is written and schemas are derived.
The teaching text lives in docs/extending.md, rewritten around Part L,
where it is read without running anything; the manifest schema, the seam
contracts and the role kinds come from the definitions themselves, because
a transcribed schema lies at the first new seam — silently, to exactly the
person who has no other source. F.203 adds a seam at runtime and asserts the
served document changes. GET /v1/extensions/authoring is deliberately
not under the admin gate: writing is not installing, an extension is
written on a laptop and installed by whoever governs the deployment, and
gating the documentation on that authority withholds it from the only
person who needs it — J.5.12's reasoning about skills, applied again.

A patch may break you, and the version now says so. Nothing in this
repository said what a version number means. The release workflow enforces
that the tag equals version in pyproject.toml, which is a consistency
check and not a policy: it never said which number to pick, and the rule
followed for thirty releases lived in one person's head. CONTRIBUTING.md
now carries it — the minor is the spec version a release implements, the
patch is every release that cuts no spec, the major stays 0 deliberately,
and while it is 0 a patch may change behaviour.

That last sentence had one consequence to chase through the tree: a
station_compat of >=0.79,<1.0 then protects nobody, and Part L had just
sold that range as a guarantee. So L.1 states what the host's version
promises and what it does not, docs/extending.md teaches the pin, the
L.16 reference prints the exact range at the one place an author copies it
from, and extensions/whisper pins to >=0.81,<0.82 — the extension
this project ships pins the way it tells everyone else to. Its suite reads
monkeyllm.__version__ rather than a literal, so the range is actually
exercised and a failure at the next minor is the pin working.

Also in this release

• The Extensions console gains the author's half. A file picker beside
  the source field, and a Write an extension panel that lists the seams
  read from this Station and hands out a folder (SKILL.md plus the
  generated references) for a coding agent. api.js now exports one base64
  encoder, so ingest and the extension upload no longer carry two.
• tests/test_version_agreement.py pins what the release guard cannot
  see: the number is written twice, in pyproject.toml and in
  monkeyllm.__version__, and only the first is compared to the tag. It
  also checks that the minor names a spec file that exists, because a
  release whose minor names no spec is a release nobody can check against a
  contract.
• A v0.80 fixture of our own failed the new contract check — an events
  handler declared echo(value), a signature no real events call could
  fill, which passed a whole release because nothing checked.

Leftovers, named

• F.194 (signed-tag verification against the forge's published keys) is
  implemented and not measured — it needs a signed tag and gpg on the
  runner.
• The curated index is still not published, so the verified tier has
  nothing to resolve against.
• The Extensions console has had no real-browser pass, the authoring panel
  included.
• F.169's curator measurement is still pending from 0.75.

Acceptance

F.199 – F.205. Full suite green locally on the release commit — 1842
collected, 0 failures, exit 0 on both matrix legs (Python 3.11.15 and
3.12.1) — and running in CI on this PR. python -m build + twine check
passed (wheel 246 kB, sdist 561 kB; the wheel carries
monkeyllm/extensions/contracts.py, authoring.py and cli_ext.py, and
nothing of the Station).