test(terminal): prove the WS ownership refusal through the real route (#14960)

[mrveiss]

Thinking Path

Closure audit of #14960 found four of its five acceptance criteria discharged by merged code
with route-level evidence, and one — "an authenticated user connecting to another user's
session id is rejected" — covered only at handler level.

api/terminal_websocket_auth_test.py::TestTerminalWebsocketOwnership::test_non_owner_is_rejected
awaits terminal_websocket(ws, session_id) directly. That style of test cannot observe a
routing or dependency-resolution failure — which is precisely how #14998 hid: the same suite
stayed green for the entire period every real handshake 500'd inside solve_dependencies,
because a router-level Request-typed Depends cannot resolve in a WebSocket scope.

A rejection criterion is about what a client experiences, so it needs a test that reaches
the guard the way a client does. #15085 added exactly that shape for the authenticated and the
unauthenticated cases; the non-owner case was the one left behind. This adds it, and nothing
else — no production code changes, no change to authentication or authorization semantics.

What Changed

autobot-backend/api/terminal_websocket_route_test.py only.

• New TestTerminalWebsocketRouteHandshake::test_authenticated_non_owner_is_refused: a real
  TestClient handshake on /ws/{session_id} as mallory against a session owned by alice.
  Asserts close code 1008 observed on the wire as a WebSocketDisconnect, that accept()
  never ran (so no terminal handler is constructed and no shell starts), and — via caplog —
  that the ownership branch fired, not the unknown-session branch of Unknown terminal session_id defaults to a live STANDARD-security terminal #14961. Both branches
  close 1008 from the same line, so without the log assertion the test would pass for the
  wrong reason.
• WebSocketDisconnect hoisted to module scope and the two existing function-local imports
  dropped in favour of it. This is what keeps the file inside the 600-line cap: 599 lines,
  up from 584 minus the 4 lines the hoist recovers.

api/terminal.py is untouched and unchanged in size at its grandfathered 1299 lines.

Verification

Targeted run, this file only:

api/terminal_websocket_route_test.py ...................  [100%]
19 passed in 4.71s

Contrast mutation — the point of the change. Replacing the ownership comparison in
_lookup_terminal_session with a never-taken branch:

    owner = config.get("owner")
-   if not owner or owner != user.get("username"):
+   if False:  # MUTATION: ownership check removed

1 failed, 18 passed
FAILED ...::TestTerminalWebsocketRouteHandshake::test_authenticated_non_owner_is_refused
E   Failed: DID NOT RAISE <class 'starlette.websockets.WebSocketDisconnect'>
------------------------------ Captured log call ----------------------------
ERROR    services.simple_pty: Error in PTY read loop: [Errno 9] Bad file descriptor

Exactly one test fails, the new one, and it fails for the right reason: with the check removed
the non-owner's handshake is accepted and a real PTY is allocated for them — the captured
simple_pty error is that shell being torn down. The other 18 route tests, including both
pre-existing handshake tests, pass throughout: they could not see this.

Mutation reverted before commit; git diff --stat on the branch shows one file, +19/-4.

Local fastapi is 0.135.2 against CI's 0.141.1 (#15091). The assertion here is behavioural —
a real handshake — so it is unaffected by the deferred-include_router differences that made
introspection unusable in #14998; the file's structural checks already run through a
clean-interpreter subprocess dump and are untouched.

Model Used

Claude Opus 5 (1M context)

Refs #14960.

[github-actions]

✅ SSOT Configuration Compliance: Passing

🎉 No new hardcoded values of either class — ssot and other both block.

Known backlog in pipeline-scripts/hardcoded_values_baseline.txt is suppressed and tracked in #14371.