[Taure]

Split out of #45, which is now documentation only. This is the behaviour half:
a wire-format fix, so it carries a fix: title and cuts its own release.

The bug

WorldInputAsync wrapped the caller's JSON in WsMatchInputPayload, so a
world.input frame went out as:

{"type":"world.input","seq":412,"payload":{"data":"{\"kind\":\"move\",\"x\":600}"}}

The server's world.input clause (src/ws/asobi_ws_handler.erl) takes the
payload verbatim as the input map:

InputData =
    case maps:get(~"data", Payload, undefined) of
        undefined when is_map(Payload) -> Payload;
        Other when is_map(Other) -> Other;
        _ -> #{}
    end,

An inner data is only honoured when it is itself an object. Unlike
match.input, which JSON-decodes a data string, world.input has no such
clause, so a data string falls to _ -> #{}. Every world input this SDK
sent reached handle_input/3 as an empty map.

It failed silently, in the worst way for anyone building prediction on it:
asobi_zone:apply_inputs/4 records the ack before it calls handle_input, so
world.ack kept advancing while nothing in the zone moved.

The fix

WorldInputAsync now sends the caller's JSON as the payload, and the decision
about what that payload is lives in WsFrame.WorldInputPayload:

• null, empty or blank becomes {}, so the frame stays valid JSON
• an object goes out byte for byte, never wrapped or re-encoded
• anything else throws ArgumentException

That last one is new. The old guard was IsNullOrEmpty, which let " " and
"not json" splice onto the wire as a malformed frame, and "[1,2,3]" and
"null" reach the server as valid JSON it rejects with a badmap. The
JsonUtility.ToJson path this replaced was structurally immune to all four.
world.input is fire-and-forget with no cid, so the server has nothing to
correlate a rejection to and none of it could ever be reported back: the game
would just fail silently for its whole life. Failing loudly on the developer's
first frame is the only place the mistake is visible. Checked by first
non-whitespace character rather than parsed, because this is a per-input hot
path.

Coverage

The outbound envelope text lives in WsFrame, which has no UnityEngine
dependency, so the wire shape is asserted by the licence-free .NET tests rather
than only by the Unity smoke run. SendAsync's cid-carrying envelope was still
hand-built inline, which is most of the 24 send methods; it now goes through
WsFrame.Request and is pinned too.

125 tests pass (dotnet test Tests/AsobiCore.NET/Asobi.Core.Tests.csproj),
15 of them new:

• cid and seq are siblings of payload, never fields inside it, and seq
  is omitted entirely when unstamped
• an absent world.input is an empty map; an object is verbatim; a top-level
  data key is still sent verbatim rather than quietly rewritten
• "not json", "[1,2,3]", " [1,2,3]", "null", "42" and "\"move\""
  are all rejected

Mutation-checked: reverting WorldInputPayload to the old wrapping body fails
13 of them, and dropping cid from WsFrame.Request fails 2. The earlier
revision of this PR was green under the same revert.

Known gap. Reverting WorldInputAsync itself, rather than the decision it
delegates to, still keeps the suite green: AsobiRealtime.cs depends on
UnityEngine.JsonUtility and cannot be linked into the .NET project at all, so
no headless test can see the one-line call. The decision is pinned; the
delegation to it is not. Closing that needs a seam over ClientWebSocket so a
test can read the frame that was written, which is a bigger change than this
PR. Recorded in Tests/AsobiCore.NET/README.md next to the reconnect-loop gap.

The reserved key

data is reserved at the top level of a world.input payload: the server
unwraps it when it is a map, dropping every sibling key, and discards the whole
input when it is not (asobi_ws_handler.erl:849-854). Tracked as
widgrensit/asobi#478.

WorldInputAsync's // comment block is now /// <summary> XML docs so that
contract reaches IntelliSense. This is the one method where a Unity consumer
most needs the rules at the call site: the payload is passed through untouched,
so every constraint on it is the caller's to respect.

Source-breaking

The parameter rename data -> inputJson breaks callers that pass it by name
(WorldInputAsync(data: json)). That is a compile error, not a silent change,
so it surfaces on the next build rather than at runtime.

Also here

The world.ack comments on WsWorldAckPayload, DispatcherTests and
DispatchTests said the two payload fields deserialise directly. They do, but
OnWorldAck hands over the raw envelope, so payload has to be pulled out
before JsonUtility sees it, or you silently get tick=0, seq=0. Corrected in
place. These are comment-only edits, kept here rather than in #45 so that #45
stays a pure **.md diff and does not trip dispatch.yml or release.yml.

Not merging

Opened for review only.