Fix the defects the agents page turned up

[AlexeyShalaev]

Writing docs/agents.md against the source turned up five things. Four were real; one was
half real, and the half that was real ran the opposite way from the report. Every finding is
below with its reproduction, including the parts that did not hold up.

Nothing here is breaking. One name leaves a submodule's __all__ — see finding 5.

1. The health page said the factory does not forward options or compression

It forwards both, on purpose, with the reason in a comment right there in factory.py:
the probes have to negotiate HTTP/2 the way the application channels do. The page told the
reader the opposite and sent them off to construct HealthChecker by hand for something they
already had.

Reproduction — a settings object carrying channel options, with the checker class replaced by
a spy:

service = 'users.v1.Users'
options = [('grpc.max_receive_message_length', 1024)]
compression = None
insecure = True
check_interval = 7.0
timeout = 2.0

The doc drifted, not the code. Corrected, and the paragraph now lists what the factory does
pass, including the service the next finding is about. The "build it yourself" advice stays,
attached to the things settings genuinely cannot express — on_status_change, max_backoff,
fail_fast_callback.

While in that table: max_backoff was listed under "what the factory wires for you" without
saying that it is the checker's own default and is not read from settings. It says so now.

2. The same page on the probed service, and on where the error lives

Two claims, both wrong.

"Each probe asks for the overall server status ... rather than a per-service one."
HealthChecker.__init__ takes service: str = "", and the factory forwards
getattr(settings.health_checker, "service", "") — visible in the same reproduction above,
which came back with service = 'users.v1.Users'. The empty name is the default, not the only
option. The page now says so, and the settings table gained a service row.

"HealthCheckerNotRunningError lives in grpc_client_kit.health, which likewise needs the
extra." It lives in grpc_client_kit.errors, is exported from the package root, and its own
docstring explains that it is placed there so catching it does not require the [health]
extra: the caller who meets it is a caller of a balancer, and a balancer works on a bare
install. The page was telling people to install an extra to write an except clause.

3. Two pages said there is no settings block for their layer

resilience.md on wait-for-ready and deadlines.md on deadline budgets both said "there is
no settings block for this layer" and routed the reader to build_interceptors and a
hand-built chain. configuration.md documents both blocks correctly, and factory.py reads
both with getattr and builds the layers into the chain itself.

Reproduction — a settings object with a wait_for_ready block and a deadline_budget block,
handed to the factory, then the chain the factory built (duplicates are the four RPC-kind
adapters per logical interceptor):

AsyncTimeoutInterceptor x4
AsyncDeadlineBudgetInterceptor x4
AsyncWaitForReadyInterceptor x4
AsyncRetryInterceptor x4
AsyncCircuitBreakerInterceptor x4

Both layers are there, in the positions the pages themselves describe. Corrected. The part of
the advice that was right — per-method caps need a hand-built chain, and that chain goes to
GrpcClient(interceptors=...) and not to create_client(interceptors=...) — survives, now
attached to the thing that is actually true of it.

docs/agents.md carried the same claim in the sentence introducing its hand-built chain
example ("which is what you need for per-method budgets, request-budget propagation or
wait-for-ready"); that line is fixed in the same commit.

4. hasattr(grpc_client_kit, "HealthChecker") raises instead of returning False

Reproduces on an install without the [health] extra:

ImportError: Install grpc-client-kit[health] (grpcio-health-checking) to use gRPC health checking
  File "grpc_client_kit/__init__.py", line 164, in __getattr__

hasattr swallows AttributeError only, and the lazy __getattr__ raises ImportError.

Kept as it is. A missing extra is an install problem and should say so; downgrading to
AttributeError would make a broken install indistinguishable from a name that never existed,
and would break except ImportError around the attribute. The obvious third option does not
exist — an exception inheriting both raises TypeError: multiple bases have instance lay-out conflict, which I checked rather than assumed.

So the defect is that this was written down nowhere. It now appears in the health guide, in
rule 19 of the agents page and in the __getattr__ docstring, each naming the supported probe
(importlib.util.find_spec("grpc_health"), or catch the ImportError). The existing
bare-install subprocess probe in tests/unit/conftest.py gained an assertion for it, so a
future change to AttributeError fails a test instead of silently changing an unwritten
contract. Verified by making that change locally: the probe fails, as intended.

5. Exports that disagreed with each other

The report named the three extras protocols and MethodCircuitState. Checking the whole
surface found the same class of problem running the other way, and a third case that is not a
problem at all.

in protocols.__all__ but not at root: FullGrpcClientSettingsProtocol,
                                      GrpcChannelExtrasProtocol,
                                      GrpcObservabilityExtrasProtocol
root protocol exports missing from protocols.__all__: CircuitBreakerMetricsProtocol,
                                                      RetryMetricsProtocol
circuit_breaker.__all__ not at root: MethodCircuitState

The three extras protocols are now exported from the package root. Eleven of the module's
fourteen protocols already were; the three left out are exactly the ones describing the
optional settings blocks — which is to say, the ones a settings-object author reaches for
after the required surface. There is no principle separating them from the eleven, so the rule
is now simply "every protocol in protocols.__all__ is a top-level export". Purely additive.

RetryMetricsProtocol and CircuitBreakerMetricsProtocol are now in protocols.__all__.
They are top-level exports and documented as exported, but a star import of the module skipped
them.

MethodCircuitState leaves circuit_breaker.__all__. This is the one direction that is
not additive, so: it is the breaker's mutable per-method bookkeeping, it appears in no public
signature (get_states() returns CircuitBreakerStatus snapshots), and it is documented
nowhere. Declaring it public froze the breaker's implementation into the compatibility
contract for no caller's benefit — the same reason ChannelWrapper and chain_token are kept
out of the pool's surface, which the package comment already states. It stays importable by
name; only import * from that module changes, and the test suite imports it directly and
still passes.

Not a defect: the module-only names generally. A sweep found fifteen names in submodule
__all__s that the root does not export — Continuation, RpcType, DEFAULT_RETRYABLE_CODES,
DEFAULT_SENSITIVE_HEADERS, HAS_TRACING, the health module's constants, and so on. The
curated root is deliberate and the agents page maps the ones a caller reaches for. The three
protocols were the odd ones out, not the pattern.

Three tests in tests/unit/test_init.py pin all of this; all three fail on master:

declared in grpc_client_kit.protocols but not exported: ['FullGrpcClientSettingsProtocol', 'GrpcChannelExtrasProtocol', 'GrpcObservabilityExtrasProtocol']
exported by the package but absent from protocols.__all__: ['CircuitBreakerMetricsProtocol', 'RetryMetricsProtocol']
assert 'MethodCircuitState' not in [..., 'MethodCircuitState']

The agents page

Updated in this PR, as CONTRIBUTING.md requires: the protocols list gained the three new
exports, the module table lost their row, the internals paragraph gained
MethodCircuitState, rule 19 gained the hasattr caveat, and the hand-built-chain sentence
lost the claim from finding 3.

Verification

make check
  ruff check .            All checks passed!
  ruff format --check .   109 files already formatted
  mypy grpc_client_kit    Success: no issues found in 25 source files

make test-unit            418 passed, 92 deselected
make test-integration      90 passed, 418 deselected, 2 xfailed
make test                 508 passed, 2 xfailed — total coverage 96.02%

uv sync --frozen --all-extras --group dev; uv.lock unchanged.