feat(client): connection resilience - jittered reconnect backoff, idempotency-safe replay, and fault visibility

[st0o0]

Summary

Hardens client (and one server) connection resilience so transient failures are handled predictably instead of turning into busy-loops, silent data loss, duplicate side effects, or hung callers. Five related changes:

1. Shared exponential reconnect backoff with jitter across all protocols
2. Idempotency-safe replay + deterministic teardown for H1.1 / H2 / H3
3. Surface content-decode failures instead of silently truncating
4. Surface application-handler crashes in the server bridge
5. Test/dependency housekeeping

Details

Reconnect backoff (feat)

New ReconnectBackoff.Compute centralizes exponential-backoff-with-jitter so HTTP/1.x (ReconnectPolicy), HTTP/2 & HTTP/3 state machines, and ClientStreamOwner all space their retries identically. The first attempt after a real disconnect still fires immediately; only subsequent retries are deferred behind a reconnect-backoff timer. Spacing avoids hammering a connection-refused peer in a tight loop; jitter avoids a thundering herd of correlated reconnects.

New, fully-defaulted knobs on the client options:

• ReconnectInitialBackoff (default 100 ms), ReconnectMaxBackoff (5 s), ReconnectBackoffMultiplier (2.0), ReconnectBackoffJitter (±20%)
• StreamRetryBackoffJitter (±20%) for stream re-materialization

Behavior note: reconnect retries are now spaced by default (100 ms initial). Set ReconnectInitialBackoff = TimeSpan.Zero to restore the previous zero-delay immediate-reconnect behavior.

Idempotency-safe replay & deterministic teardown (fix)

• After connection loss, only idempotent in-flight requests are replayed. Non-idempotent requests (e.g. POST) are **failed rather than replayed*ve processed them, so a replay risks aduplicate side effect (RFC 9110 §9.2.2). Applied consistently to H1.1, H2, and H3.
• On stage teardown (KillSwitch abort → PostStop → Cleanup), requests still in flight or buffered for replay arenow failed, not silently dropped, so calf hanging until their own client-sidetimeout. request.Fail is idempotent, so streams that already delivered a response are unaffected.
  ### Content-decode failures (fix)
  A corrupt or truncated compressed response body (Content-Encoding: gzip/…) now throws a clear HttpRequestException ("compressed data is corrupt or truncated") instead of silently returning a truncated body. Typed as HttpRequestException so HttpContent's stream-copy wrapping passes it through unwrapped.

Server handler crashes (fix) ApplicationBridgeStage no longer lets ann vanish: it reaches context disposal(OnDispose(context, exception)), emits a 500, and is traced — so operators can observe handler crashes. Crashes during context creation (before an `AppConed via tracing.

Housekeeping (chore)

Removed the obsolete WireBuffer wrapper-pool test hook and bumped servus.akka.

Testing

New/expanded specs cover each change: recoolicy scheduling,idempotent-vs-non-idempotent replay, teardown-fails-buffered/in-flight requests across H1.1/H2/H3, corrupt-gzip decode surfacing, and handler-crash disposeen.