nx_mcp_stdio_samepool_gate_t161.nx

nx_mcp_stdio_gate.nx -- lane conn, 2026-09-11. PROVES the connector's per-request failure semantics with REAL child processes, in-process (the connector's main-bearing file cannot be imported, so the recovery + emit logic it drives lives in nx_mcp_pending, which this gate imports directly): * a worker that dies WITHOUT self-reporting (a crash, or the SIGALRM deadline that kills it) becomes a CORRELATED JSON-RPC error for THAT id and the loop CONTINUES (mp_recover_pump returns >= 0); * the emitted error carries the RIGHT id and the JSON-RPC error shape; * a LATER request still succeeds after the failure; * the deadline (death by SIGALRM) maps to the -32012 deadline error, not a generic one; * mo_prefix_id extracts a top-level id from a truncated request prefix (string, number, truncated, nested-only) -- the input to the oversized/malformed -32600 path; * a NEG-CONTROL: a SUCCESSFUL worker is forwarded verbatim and NEVER decorated with an error. BITE-PROVED MANUALLY (recorded in the lane transcript, not via a forked subject elf since the subject is in-process): reverting the failed_slot mechanism so mp_recover_pump returns -1 on any rp_pump -1 makes failing-child-continues + later-request-succeeds go RED; restoring the fix returns them GREEN. license_tier: ORIGINAL No hw writes (Rule 26).