feat(health): return structured probe outcomes

[SunSi12138]

Closes #655

Summary

• add SharpLinkHealthProbeOutcome with Success, NotReady, Unavailable, and Unsupported
• keep remote Ready / Draining / Unhealthy in nullable SharpLinkHealthCheckResult.Status, populated only for Success
• make zero-ready, in-flight remote connection loss/deadline, and missing HealthCheck capability normal structured results
• preserve caller cancellation, canonical terminal lifecycle, protocol corruption, programmer/configuration error, and unexpected runtime failures as exceptions
• keep protocol v2 health wire grammar unchanged and avoid ordinary RPC endpoint-selection/hot-path changes
• map structured results explicitly in SharpLinkRemoteHealthCheck; Unsupported has a documented stable Unhealthy mapping
• document the result/exception/Hosting contract and [Runtime Architecture][Phase 18] 完成 public API/package 审计、发布门禁与迁移文档 #86 migration note

Validation coverage

• remote Ready / Draining / Unhealthy => Success + remote status
• zero Ready connection => NotReady, with no connection side effect
• hosted Client Running + NotReady => structured NotReady
• LifecycleState=Running while legacy connection State=Faulted => structured NotReady
• remote connection closes during probe => Unavailable
• local StopAsync races an in-flight probe => terminal ConnectionClosed remains exceptional
• peer does not negotiate HealthCheck => Unsupported and no health frame is sent
• caller cancellation => OperationCanceledException
• malformed HealthResponse => SharpLinkErrorCode.ProtocolViolation
• multi-cluster scoped health preserves the same structured result contract
• Hosting maps all six normal result combinations without attaching expected exceptions

Pre-PR self-check

• rebuilt on latest dev after feat(api): add non-throwing multi-cluster status query #664 before opening the PR; current branch is ahead 5 / behind 0
• final diff is limited to 6 files: health result, health query, Hosting mapping, focused tests, and health contract documentation
• checked namespace/type references against current dev and existing test helpers/fakes
• checked indentation/braces and nullable Status call sites touched by this change; CI formatting gate passes
• Release build passes with 0 warnings and 0 errors
• HealthResponseCodec refuses to serialize non-success local outcomes, so local outcome values cannot leak onto the protocol wire
• no RpcChannel, pending-call terminal machinery, retry/admission, or ordinary RPC invocation files changed

Acceptance self-review

• structured result represents remote status plus local NotReady / Unavailable / Unsupported
• local probe outcomes never masquerade as remote Unhealthy
• single-cluster and multi-cluster scoped health use the same contract
• caller cancellation remains cancellation
• malformed/protocol-invalid response remains a protocol failure
• canonical local terminal lifecycle remains exceptional before and during a probe
• Hosting uses structured mapping for normal states instead of broad exception handling
• health query does not reconnect or make the Client Ready
• Protocol v2 wire format and ordinary RPC hot path remain unchanged
• XML/result documentation and dedicated migration/contract documentation added

CI follow-up

Two initial PR Fast failures were followed to root cause and fixed before first review readiness. Both were setup defects in the new multi-cluster test, not production-code failures:

1. the test omitted the required explicit multi-cluster request-timeout policy; fixed with DisableRequestTimeout()
2. the test used a slot without a static route; fixed with the repository-standard AllowDynamicContracts dynamic-only slot configuration

Review follow-up

Review found two lifecycle-domain races after #658:

1. legacy connection State == Faulted can coexist with canonical LifecycleState == Running; health preflight now uses LifecycleState, so zero-ready connectivity fault remains structured NotReady
2. local shutdown can close an already-selected probe connection; expected-unavailability mapping now re-checks canonical terminal lifecycle and rethrows the winning terminal exception instead of returning Unavailable

Regressions freeze the actual Running / connection Faulted / NotReady supervisor window and race an in-flight health probe against local StopAsync.

Latest exact head 4b1feb2d1cdfbb2c2e69916194f021c4c6f0d8d3 is green on:

• PR Fast — success, including formatting, Release build, Unit Tests, generator/load tests, and allocation gate
• PR Quick — success, including Debug/Release builds, Unit/Integration tests, NativeAOT, pack/NuGet contract + smoke, load and chaos smoke
• Pending and codec validation — success
• Extension fault containment — success
• CodeQL — success

Both review threads were replied to and resolved. Ready for review.

[SunSi12138]

整体 result/Hosting 方向符合 #655，但发现 1 个状态域 blocker：health preflight 仍把 legacy connection State.Faulted 当成 terminal lifecycle，导致合法 Running+NotReady 窗口异常泄漏。

[SunSi12138]

补充同一 lifecycle/result 边界的另一处 race：probe 启动后本地 Stop/terminal shutdown 会被当前 catch 重新分类为普通 Unavailable，与 PR 文档声称的 terminal lifecycle exceptional contract 不一致。