stdio_client crashes on malformed UTF-8 from child stdout instead of surfacing parse error

Comece em mcp.client.stdio.stdio_client() e tests/client/test_stdio.py; revise como a decodificação de TextReceiveStream e a iteração do fluxo de leitura lidam com erros de codificação. Adicione a regressão malformado-e-depois-válido descrita na issue e execute o teste focado para verificar que o primeiro item é uma Exception, o segundo é uma SessionMessage válida e que o transporte permanece ativo.

Summary

mcp.client.stdio.stdio_client() crashes when the spawned child process writes invalid UTF-8 bytes to stdout.

The transport currently decodes child stdout with encoding_error_handler="strict", so malformed bytes raise during TextReceiveStream(...) iteration. That exception escapes the decoding loop and brings down the transport task group instead of surfacing the bad line as an in-stream parse error.

Why this looks like a bug

The SDK already hardened the server side for the analogous case in PR #2302 (fix: handle non-UTF-8 bytes in stdio server stdin). That change explicitly preferred:

• replace invalid bytes with U+FFFD,
• let JSON validation fail on the malformed line, and
• keep the transport alive so subsequent valid messages can still be processed.

The client side still behaves asymmetrically today. A buggy or non-compliant child server can kill the Python client transport with a single malformed line even if the next line is valid JSON-RPC.

That seems inconsistent with the current stdio robustness direction.

Reproduction

A minimal child process that writes one malformed line and then one valid JSON-RPC line:

import sys
import time

sys.stdout.buffer.write(b"\xff\xfe\n")
sys.stdout.buffer.write(b'{"jsonrpc":"2.0","id":1,"method":"ping"}\n')
sys.stdout.buffer.flush()
time.sleep(0.2)

With current stdio_client(...) defaults, the transport raises ExceptionGroup instead of continuing.

Expected behavior

The malformed line should be surfaced as an in-stream parse / validation error, and the next valid JSON-RPC line should still be received.

Observed behavior

The transport task group fails before the valid follow-up message is delivered.

Proposed fix

Match the server-side approach from PR #2302:

1. default StdioServerParameters.encoding_error_handler to "replace"
2. continue treating malformed decoded lines as JSON validation failures
3. keep the background stdio tasks resilient during early subprocess shutdown / abrupt close

Validation

I reproduced this locally against current main and verified that a minimal patch plus a regression test fixes it.

A focused regression test in tests/client/test_stdio.py can assert:

• first item from the read stream is an Exception
• second item is the valid SessionMessage