feat(state)!: remove Kind; a name is the contract

[ciaransweet]

Kind was a nominal type two toolsets had to spell identically. Between
toolsets whose authors already talk to each other it worked; anywhere else it
went mute, and the way it went mute was the problem. _bindable dropped an
unmatched declaration, the parameter returned to the model's schema, and the
model invented a value. The misconfiguration and the invented geometry were
the same silent event.

What replaces it is the name a value is already stored under. Every
ToolResult data key is captured to <toolset>/<tool>/<field>, the model
reads that key off a [state updated: ...] breadcrumb and passes
@state:<key>, and the client substitutes the payload on the way out.
Producer and consumer still share nothing else — but a name degrades to a
model reading it, where a kind degraded to nothing at all. The tool segment is
in the key because the key is now read: gazet/get_aoi/bbox and
gazet/generate_aoi/bbox are different values, and under one shared BBOX
kind a consumer could not say which it wanted.

NotAuthored

NotAuthored covers what model_generatable=False covered, without the
coupling. It is a claim about one parameter — this caller must supply a value
that already exists — naming no type, so a tool carrying it works against a
client that has never heard of any of this.

What a client does with it scales with what the client implements. One that
ignores _meta leaves the parameter alone. One that reads the description
finds a sentence appended saying the value must already exist. This client
narrows the schema to {"type": "string", "pattern": "^@state:"}, so a
literal does not type-check and the model has no way to express an invented
value.

A call that writes one anyway, or omits a required one, is refused with a
listing of what session state actually holds. Refusals come back as tool
results, never exceptions: raising would leave an assistant message whose
tool_calls no ToolMessage answers, which most providers reject outright,
and one refusal would poison every later turn on the thread.

Provenance

StateEntry.inputs records, per producing call, where each argument came
from: another state key, or model. Recorded and never enforced, one level
deep, and the input side only — a tool owns its outputs, and an equality test
on returns would catch an echo while missing every transformation.

Verified against a real consumer

Ported ecmwf/dss-agentic-ai-services onto this branch and ran it, which is
where the last four commits come from. Two findings worth stating:

A model reading a dataset schema learned that CDS's own request body takes an
area, read the published bounding box back with inspect_state, and filled
both the opaque request dict and the narrowed parameter. Nothing in the
mechanism failed — the schema was narrowed, the handle was dereferenced, the
receipt was minted — but the marker binds a parameter, not a concept, and
a tool with an opaque sibling accepting the same value voids it. That limit is
now stated on NotAuthored itself, where an author will read it before
relying on it.

Worse, and now fixed: the refusal for a handle nested inside an opaque
argument ended "read a value with inspect_state and write the field yourself".
Right for an ordinary opaque field, and an instruction to carry the value
around the constraint on a tool that declares one. The model obeyed it. A
refusal is a prompt, and it has to be checked against every constraint the
same tool declares.

Examples

The client is where this is legible, so it was rebuilt rather than merely
ported: the state panel carries the whole of an entry's inputs, and hovering
any of a key, a call or an activity lights the same three-sided link. Four
rendering faults went with it, every one found by looking at the running UI —
worth noting that none of examples/agui-events/web is covered by a CI job,
which is why.

Not included

No migration guide: we are the only consumers and both ports are done or in
hand. Two known gaps are recorded rather than fixed — the listing note is a
judgement about that surface rather than about the record, and describe's
shape summary says nothing about what a value means, deliberately.

BREAKING CHANGE: Kind, mcp_runtime.kinds, mcp_state.wiring, detect_kind, entries_of_kind, published_kinds, publishers, BY_DECLARATION, BY_HANDLE, supplied, offers_handles and BuiltAgent.withheld are removed, along with the tools.withheld AG-UI activity and kind on every state entry the API serves. offer_handles loses its skip argument. State keys gain the producing tool, so a two-part key is now three parts.

[yellowcap]

This makes a lot of sense and does not degrade performance. For testing, I did make a port of the geo-assistant to this and it worked really well. Thanks @ciaransweet !

[ciaransweet]

This makes a lot of sense and does not degrade performance. For testing, I did make a port of the geo-assistant to this and it worked really well. Thanks @ciaransweet !

Amazing! Thanks!