feat(live): add ZIP_STORED visual frame pack and unpack helpers

Live video reaches ADK as `image/*` frames, alongside `audio/*` chunks. Saving each frame as its own artifact costs one storage operation per frame, while adding a batch method to `BaseArtifactService` would grow the abstract storage interface for every backend.

This change adds `google.adk.live._visual_frames`, which packs a frame sequence into one uncompressed (`zipfile.ZIP_STORED`) `.zip` archive that the existing `save_artifact()` / `load_artifact()` methods store as is:

- `VisualFrame`: a Pydantic model pairing a frame's `types.Blob` with its capture timestamp.
- `pack_visual_frames()`: validates the frames (non-empty data, finite and non-decreasing timestamps), writes `frames/frame_NNNN.<ext>` members and a `metadata.json` timing manifest, and returns `(archive_bytes, manifest)`.
- `unpack_visual_frames()`: the inverse; restores `(list[VisualFrame], manifest)` from the archive bytes.
- `summarize_manifest()`: the manifest without its per-frame index, small enough for an artifact's `custom_metadata`.
- `VISUAL_FRAMES_ZIP_MIME_TYPE`: the archive's MIME type, `application/zip`.

NO_UNIT_GUIDE=Internal pack/unpack helper module for live visual frames
PiperOrigin-RevId: 982888971

Author: google-genai-bot

src/google/adk/live/_live_llm_flow.py

@@ -155,15 +155,9 @@ async def send_to_model(
           types.LiveClientRealtimeInput(audio_stream_end=True)  # type: ignore[arg-type]
       )
     elif live_request.blob:
-      # Cache input audio chunks before flushing. The cache concatenates
-      # every chunk into one audio file, so other blobs (e.g. video frames)
-      # must stay out of it.
-      if (
-          run_config.save_live_blob
-          and live_request.blob.mime_type
-          and live_request.blob.mime_type.startswith('audio/')
-      ):
-        realtime_cache_manager.cache_audio(
+      # Cache input audio chunks before flushing.
+      if run_config.save_live_blob:
+        realtime_cache_manager.cache_blob(
             invocation_context, live_request.blob, cache_type='input'
         )
 
@@ -316,24 +310,15 @@ def get_author_for_event(llm_response: LlmResponse) -> str:
         ) as postprocess_agen:
           async for event in postprocess_agen:
             # Cache output audio chunks from model responses
-            # TODO: support video data
             if (
                 run_config.save_live_blob
                 and event.content
                 and event.content.parts
             ):
               for part in event.content.parts:
-                if (
-                    part.inline_data
-                    and part.inline_data.mime_type
-                    and part.inline_data.mime_type.startswith('audio/')
-                ):
-                  audio_blob = types.Blob(
-                      data=part.inline_data.data,
-                      mime_type=part.inline_data.mime_type,
-                  )
-                  realtime_cache_manager.cache_audio(
-                      invocation_context, audio_blob, cache_type='output'
+                if part.inline_data:
+                  realtime_cache_manager.cache_blob(
+                      invocation_context, part.inline_data, cache_type='output'
                   )
 
             yield event

src/google/adk/live/_realtime_cache_manager.py

@@ -119,6 +119,35 @@ def cache_audio(
         len(cache),
     )
 
+  def cache_blob(
+      self,
+      invocation_context: InvocationContext,
+      blob: types.Blob,
+      cache_type: str,
+  ) -> None:
+    """Cache an incoming user or outgoing model blob if it is audio.
+
+    The audio cache concatenates every chunk into one audio file, so other
+    blobs (e.g. video frames) are skipped. Audio blobs are cached as copies,
+    so later changes to the caller's blob do not change the cache.
+
+    Args:
+      invocation_context: The current invocation context.
+      blob: The blob to cache.
+      cache_type: Type of data to cache, either 'input' or 'output'.
+
+    Raises:
+      ValueError: If the blob is audio and cache_type is not 'input' or
+        'output'.
+    """
+    # TODO: cache image blobs once media caching lands.
+    if blob.mime_type and blob.mime_type.startswith('audio/'):
+      self.cache_audio(
+          invocation_context,
+          types.Blob(data=blob.data, mime_type=blob.mime_type),
+          cache_type,
+      )
+
   async def flush_caches(
       self,
       invocation_context: InvocationContext,