gate: streaming — the model's events relayed as they come, the fringe as the last block

Bamidbar 15:39 u'reitem oto — the fringe must be seen, so a streaming
client gets it too. With "stream": true the request is opened
asynchronously (Gate.Upstream.open_stream/4, httpc stream: :self) and
Gate.Stream relays every complete SSE event to the caller unchanged
while gathering the text deltas; the terminal events (Anthropic
message_delta + message_stop; OpenAI the finish_reason chunk + [DONE])
are held until the gathered text has been examined, then released
behind the fringe sent as one more text block, with the `bikoret` object
added to the terminal event the client already parses. No X-Bikoret-*
headers on a stream (they leave before the model speaks); no rebuke on a
stream (nothing streamed can be taken back — 422). An examination that
fails after the stream began is said in the trailer ("— bikoret: not
examined (…)"), never silent. One api_calls_external gathered before the
call, as for any pass. Tamid §1.1, Construction §3.10, KOHEN. Doorpost
40e7d7d1…. Elixir 612/0 (read before this commit).

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

specs/bikoret_tamid.md

@@ -458,7 +458,21 @@ Response (200): the upstream's reply, with
                    rebuked?: {first_bedikah_ref, first_verdict}, disclosure }
 
   Any non-200 from the upstream is returned as it came (X-Bikoret-Verdict: none).
-  A stream request is refused (422): a fringe needs the whole garment.
+
+Streaming ("stream": true): the model's events are relayed as they
+  arrive, bytes unchanged, while the text is gathered; when the model's
+  stream ends the text is examined and the fringe is sent as ONE MORE
+  BLOCK before the terminal events — Anthropic: content_block_start /
+  content_block_delta / content_block_stop at the next index; OpenAI: one
+  more chunk with delta.content — and the `bikoret` object rides on the
+  terminal event the client already parses (message_delta; the chunk
+  with finish_reason), where an unknown field is ignored. Headers go out
+  before the model speaks, so a streamed reply has NO X-Bikoret-* headers
+  and cannot be rebuked ("stream" with X-Bikoret-Rebuke → 422). If the
+  examination fails after the stream has begun, the trailer says
+  "— bikoret: not examined (<why>)" and the object's verdict is "none":
+  delivered, marked, never silent. Gathered text is capped at 65,536
+  bytes like any output.
 
 Rebuke once: when X-Bikoret-Rebuke is "once" and the first reply is
 lo_tov, one second request = the first + the first reply as an