M4: media upload/download/playback (both directions)

Gateway plugin:
- media.upload (chunked binary) -> size/sha256 verify + MIME re-sniff ->
  cache_*_from_bytes -> media.upload.ack
- media.offer / media.pull (chunked) for agent-sent media, delivery-path
  security re-checked at pull time
- send_* overrides mint media_id and emit media.offer
- message.send media_refs resolve to cached inbound media
- per-send + per-chunk timeouts so a stalled peer can't starve the rest

App (Kotlin CMP):
- Protocol: media frame types/payloads/builders
- GatewayClient: binary session, uploadMedia (chunked + streaming sha256),
  pullMedia serialized via Mutex so concurrent offers don't interleave
- ChatStore/IrisController: MediaItem, attachments, auto-pull on offer
- Platform media: SAF picker, ExoPlayer (audio mini-player + video), image
  loader, FileProvider document open (Android); AWT-free desktop actuals
- ChatScreen: attach button + chips, media rendering, keyboard dismiss on send

UI polish:
- preserve image aspect ratio (no stretching), cap dominant dimension
- adjustResize so only chat content squeezes for the keyboard
- clear focus (hide keyboard) on send

Docs: media.upload.ack in 04-wire-protocol.md + frames.schema.json +
07-media.md; M4 marked complete in 14-milestones.md.

Tests: 17-test tests/gateway/test_android.py suite passes.

docs/07-media.md

@@ -29,9 +29,10 @@ receipt (don't trust the client) using hermes helpers
    - image → `cache_image_from_bytes`
    - audio/voice → `cache_audio_from_bytes`
    - video → `cache_video_from_bytes`
-   - document → `cache_document_from_bytes`
-   → returns a local path.
-6. The path is attached to the next `message.send` via `media_refs`, becoming
+- document → `cache_document_from_bytes`
+    → returns a local path.
+ 5b. Plugin replies `media.upload.ack {ok, media_ref}` (failures use `error`).
+ 6. The path is attached to the next `message.send` via `media_refs`, becoming
    `MessageEvent.media_urls` + `media_types`
    (`gateway/platforms/base.py:2337`). The agent's vision/audio tools can then
    read the file.