fix: splice a list-valued websocket reply into the command list

[Taure]

Closes #399.

nova_websocket:call_result/0 documents

{reply, OutFrame :: cow_ws:frame() | [OutFrame :: cow_ws:frame()], State :: map()}

but the list form killed the connection process.

nova_basic_handler:handle_ws/2 consed the payload onto commands as a single element. nova_ws_handler:handle_ws/4 hands that list to cowboy untouched, and cowboy treats every top-level element as one command — so a list-valued reply reached cow_ws:frame/2 whole:

{function_clause,[{cow_ws,frame,[[{text,<<...>>},{text,<<...>>}],#{}],
   [{file,"cow_ws.erl"},{line,875}]},
 {cowboy_websocket,commands,3,...},
 {cowboy_websocket,handler_call_result,5,...}]}

Ranch logs the connection process exiting and the client's socket drops.

The fix

prepend_frames(Frames, Commands) when is_list(Frames) -> Frames ++ Commands;
prepend_frames(Frame, Commands) -> [Frame|Commands].

Order matters and is not reversed: handle_ws/4 returns the accumulator to cowboy as-is, so its order is the order frames reach the wire. Splicing in place is therefore correct and lists:reverse/1 would be wrong.

No cow_ws:frame() is itself a list — they are atoms (close, ping, pong) and tuples — so is_list/1 separates the two shapes unambiguously. An empty list contributes nothing.

Tests

Both levels, because the defect is only visible at the lower one:

• nova_basic_handler_tests — the command list built for a single frame, a list, an empty list, the hibernate variant, order preservation, earlier commands kept, and bare atom frames.
• nova_ws_handler_tests — what cowboy actually receives from websocket_info/2 and websocket_handle/2, plus an assertion that every element survives cow_ws:frame/2. That last one is what would have caught the original bug: the controller's return value is correct in isolation, so a test one level up passes while the connection dies.

Reverting the source change fails 8 of them.

Checks

xref, dialyzer (no warnings in the changed files), eunit 356/356, ct, ex_doc. elp lint reports one pre-existing L0002 on nova_basic_handler.erl that is also present on master (an ELP internal crash, not a code defect).

How it was found

An application emitted two frames from one websocket_info/2 clause. EUnit passed — the return value is right — and Common Test failed, because the defect only exists once Nova translates the return into cowboy commands.