…action tests

Extends the interaction suite with the initialize handshake (server identity,
instructions, capability derivation, client identity and capabilities as seen
by the server), completion round trips, logging notifications, and the
MCPServer resource/prompt/structured-output behaviours. Records two more
divergences on the requirements manifest: MCPServer reports unknown resources
and prompts with error code 0 rather than the codes the spec documents.

Removes the 'pragma: no cover' from the method-not-found fallback now that it
is covered.

Covers the server-to-client half of the interaction model: sampling,
form-mode elicitation, roots, progress in both directions, list_changed
notifications, and request cancellation, all against the low-level Server
through the public Client API. Records a further divergence: the server
answers cancelled requests with an error response where the spec says no
response should be sent.

Removes five more 'pragma: no cover' comments on paths these tests now
cover (server list_changed senders, the client roots send path, and the
default elicitation callback).

…eta interaction tests

Covers URL-mode elicitation (including the elicitation/complete lifecycle and
the -32042 rejection flow), resource subscriptions and update notifications,
cursor pagination across all four list methods, request and session read
timeouts, _meta round trips, and the MCPServer Context convenience methods.

Removes the 'pragma: no cover' from the resource-updated send path now that
it is covered.

…ction tests

Adds ClientSession-level tests for pre-initialization request rejection and
protocol version negotiation, a proof that concurrent tool calls are
dispatched simultaneously and answered independently, and tests pinning
three behaviour gaps: tool-set mutations send no list_changed notification,
logging/setLevel is not supported by MCPServer and no level filtering
exists, and tool-enabled sampling is rejected because the high-level client
cannot declare the sampling.tools capability.

A RecordingTransport wrapper tees every message crossing the client's
transport boundary so the suite can assert properties that are invisible to
API callers: request ids are unique and never null, notifications are never
answered, and exactly one initialized notification is sent between the
initialize response and the first feature request.

Drives the streamable HTTP Starlette app through httpx's ASGI transport so
the full HTTP framing layer (session ids, SSE and JSON response encoding,
stateful and stateless modes) runs with no sockets, threads, or
subprocesses. Covers the handshake, tool calls and errors, mid-call
notifications, the stateless rejection of server-initiated requests, and
the routing of unrelated server messages to the standalone stream.

Removes the 'pragma: no cover' comments these tests now cover (the
session-manager accessors, the no-session-id validation path, and the
related-request-id routing branch). The session-manager accessor's
unreachable error guard keeps its pragma, moved onto the raise statement
itself so the now-executed condition above it is measured normally.

…irements manifest

Fixes the spec deep links that pointed at non-existent anchors, records the
divergences for the client's default not-supported answers (the spec names
-32601 for roots and -32602 for an undeclared elicitation mode; the default
callbacks answer -32600), and adds a logging:capability requirement noting
that MCPServer emits log message notifications without declaring the
logging capability.

Also tightens behaviour sentences and docstrings to match what the tests
assert, and adds a test pinning that Context.report_progress is a silent
no-op when the caller supplied no progress token, removing the pragma on
that path.

…n-rejection interaction tests

… two-way test coverage

…tiated requests over streamable HTTP

…TP transports

… conformance tests

…ence model

…ments

…p requirements

…ator stacking

SseServerTransport.connect_sse never closed sse_stream_reader after
EventSourceResponse returned; the in-process bridge dropped its chunk
reader when a request was cancelled before the response started. Close
both at source so the interaction suite no longer needs a gc.collect()
workaround, and pull one assert inside its async-with body to clear the
last 3.11 coverage gap.

The False arc exits through the enclosing async-with and is dropped by
coverage.py on 3.10 only; 3.11+ record both arcs.

…pass

- _provider.py docstrings now describe the provider, not its authoring history
- eight manifest behavior/divergence notes scoped to what the spec/SDK actually
  state (logging field-shape only, sse endpoint-event spec-only, scope-selection
  AS-fallback step, sampling result-balance ValueError, hosting:session:delete
  doesn't remove transport, low-level elicitation validation/restriction scope,
  progress no-token second clause)
- client-transport:http:reconnect-get deferred reason no longer claims the
  feature is unimplemented
- rename the sampling mixed-content rejection test to match what it asserts
- pagination cursor pass-through asserted by identity per the suite convention
- mcpserver:prompt:unknown-name docstring acknowledges the code-0 divergence
- test_bridge cancel_on_close test bounds the transport-close wait
- drop now-stale no-branch pragma at shared/auth.py:94

Nine notification tests sent capability-gated messages (log, resource_updated,
roots/list_changed, *_list_changed) from a peer that had not declared the
capability, which only works because the SDK does not yet enforce capability
gating. Adding stub handlers / list_roots_callback so the capability is
advertised makes the tests match their requirement preconditions and survive
the gate fix unchanged.

The three list_changed tests cannot set listChanged=True without threading
NotificationOptions through the in-memory and HTTP-manager connection paths;
the requirement behavior text now describes what the tests prove (send ->
arrives) and the module docstring records the remaining coupling on
lifecycle:capability:server-not-advertised.