feat(teams): outbound reactions, targeted ephemeral messages, placeholder-aware native streaming (#219)

CHANGELOG.md

@@ -4,6 +4,12 @@
 
 Sync wave from `chat@4.31.0` to `chat@4.41.1` (tracking #184). `UPSTREAM_PARITY` stays `4.31.0` until #203.
 
+- **Teams outbound: reactions, targeted ephemeral messages, placeholder-aware native streaming** (#219). Ports vercel/chat `5eb8b846` (#734), `160140e3` (#737) and `93a58af5` (#709) from chat@4.35.0, and `85089037` (#951, chat@4.41.0). A live Teams check is pending.
+  - **Consumer-visible (Teams):** `add_reaction` / `remove_reaction` now call the Teams conversations reactions API instead of logging a warning, and can raise (`AuthenticationError`, `AdapterPermissionError`, `NetworkError`, `AdapterRateLimitError`). `check`, `eyes`, `pin`, `rocket`, `thinking`, `thumbs_up` and `x` map to their Teams IDs; any other name is sent as a native Teams reaction ID (`like`, `heart`, `1f440_eyes`, ...).
+  - **Consumer-visible (Teams):** `TeamsAdapter.post_ephemeral` is new. In channels and group chats, `thread.post_ephemeral(user, ...)` now sends a native targeted message that only `user` sees (`used_fallback=False`), instead of opening a DM. The app must be installed in that conversation. In a 1:1 chat it posts normally with `used_fallback=True`. Edits and deletes of a targeted message go through the targeted endpoint: the sent id is recorded in the state adapter under `teams:targetedActivity:{conversationId}:{messageId}` for 24 hours (the key upstream uses), and state errors fall back to the plain endpoint.
+  - **Consumer-visible (Teams streaming):** only apps that set `fallback_streaming_placeholder_text` explicitly see a change. In a DM with a native streamer, the placeholder is sent as the Teams informative status line before the first chunk. In group chats, channels and proactive messages, `TeamsAdapter.stream()` returns `None`, so core posts the placeholder and edits it as text arrives, instead of posting one buffered message. `""` counts as set. Unset or `None` keeps the previous behaviour exactly. `TeamsAdapter.stream()` now returns `RawMessage | None`.
+  - **Python-specific (divergence from upstream):** a reaction ID that is not `[A-Za-z0-9_-]+` raises `ValidationError` before any request, because the Python SDK puts it in the URL path unescaped. `_stream_via_emit` waits at most 30 s (`STREAM_FIRST_CHUNK_ID_TIMEOUT_S`) for the first chunk's id, then settles the stream and, if nothing was delivered and the user did not cancel, posts the accumulated text as one ordinary message: on `microsoft-teams-apps` 2.0.16+, a terminal 403 on the first flush (for example "Content stream is not allowed") never delivers a chunk and never sets `canceled`, so the wait used to hang the DM handler, and the reply was never delivered. See `docs/UPSTREAM_SYNC.md`.
+  - Works on `microsoft-teams-apps` 2.0.13 to 2.1: reactions use `conversations.add_reaction` / `delete_reaction` where the SDK has them (2.0.16+) and `ApiClient.reactions` otherwise.
 - **Turn cancellation, typing lifecycle and agent-session events** (#201; core half of vercel/chat `2ce2be00` #862, chat@4.39.0). No default behavior change: state polling runs only for adapters that set `supports_turn_cancellation` (none in-repo yet; Slack lands in #215).
   - **New `thread.signal`** (a `TurnSignal`: `aborted`, `await signal.wait()`, `add_listener(cb)`), aborted when the turn is stopped. Pass it to model calls so a stop also stops generation. Streams posted from the turn stop being consumed when it is aborted: the reply is finalized with what was already streamed, and the source generator is closed. Outside a message handler the signal is never aborted.
   - **New `chat.abort_turn(thread_id)`**: aborts the thread's running turn in this process and, when the turn's adapter opted in, in any process sharing the state backend (`active-turn:` / `abort-turn:` keys, 1 h TTL, polled every 250 ms per running turn).