feat(go2): native CompressedImage across all go2 blueprints (#2831 option 5)

[spomichter]

Problem

#2831 settled on option 5: connections output CompressedImage directly, producers pick the encoding, consumers decode only when they need pixels. This PR does that migration for every go2 blueprint. Stacked on #2814 (the CompressedImage type).

What changes

Producer — GO2Connection.color_image: Out[CompressedImage]. jpeg encode happens once at the publish edge (config.image_quality, default 75). frames that are already CompressedImage pass through untouched — which is also why old replay datasets keep working with zero conversion: replay feeds raw Image frames through the same edge and they get encoded like live frames. Go2FleetConnection inherits the change.

Consumers — decode at point of use, and CompressedImage.decode() is memoized (one decode per frame max, cache never pickles):

• go2-only modules flip in place: PerceiveLoopSkill, PersonFollowSkillContainer, TemporalMemory, Go2Memory, Go2Mid360Recorder, VLMAgent, VlmStreamTester, Go2TeleopModule
• modules shared with g1/xarm/desk graphs get port-retype subclasses so those robots keep raw Image until they migrate: CompressedSpatialMemory, CompressedNavigationSkillContainer, CompressedDetection3DModule, CompressedMarkerDetectionStreamModule, HostedTwistTeleopModule (retyped in place — hosted-go2 only)
• note for review: the port retypes need # type: ignore[assignment] (deliberate LSP violation — that IS the wiring mechanism). 5 of them, each commented.

Zero-decode paths (the structural wins over the codec wrapper):

• VLMAgent sends the jpeg wire bytes straight to the model (CompressedImage.agent_encode()) — the old per-query re-encode is gone
• quest teleop _push_jpeg forwards wire bytes to the headset — per-frame re-encode gone
• CameraVideoTrack (hosted teleop webrtc) decodes at recv(), so only frames actually sent to the operator pay a decode
• rerun viewer renders the jpeg via rr.EncodedImage, no pixel decode
• recorders store CompressedImage → new datasets ~39x smaller

Misc: sensor_msgs.CompressedImage added to zenoh latest-wins QoS; bench harness gets --sink decode|bytes and per-nic wire sampling; bench blueprint drops the viewer bridge (it was hijacking any open viewer on the box).

Native e2e benchmark (same protocol as #2831: go2_short replay ~14Hz 720p, 60s x 2 reps)

cell                     frames   fps     lag  cpu%  rss MB  wire/frame
basic  clean  bytes         855  14.3   -52ms     2    1836        71KB
basic  clean  decode        854  14.3   -51ms     2    1893        71KB
go2    clean  bytes         854  14.3   -51ms     4    4732        71KB
go2    clean  decode        855  14.3   -69ms     4    4967        71KB
go2    jetson bytes         853  14.3   -55ms     2    5455        71KB
go2    jetson decode        854  14.3   -53ms     2    5528        71KB
heavy  clean  decode        855  14.3   -52ms     6    5102        71KB
heavy  jetson decode        854  14.3   -39ms     3    5878        71KB

against the #2831 baselines (same box, same dataset):

cell	raw	codec pin	native
go2 / jetson fps	6.5	14.3	14.3
heavy / jetson fps	11.4	14.3	14.3
basic / clean fps	3.3	14.3	14.3
go2 / clean cpu	17%	4%	4%
heavy / jetson cpu	13%	3%	3%
wire per frame	2.76MB	~140KB (synthetic)	71KB (real frames)

native matches the codec pin on delivery and cpu and beats raw everywhere, with no wrapper in any blueprint — the encode lives in the connection where option 5 wants it. real dataset frames compress to 71KB at q75 (39x), better than the synthetic 720p estimate.

pending: network sweep (rate-limit + loss profiles via tc netem) — needs sudo on the bench box; raw/codec network baselines run from 3f2fc05ad where the harness still has --mode raw|codec.

Breaking Changes

• GO2Connection.color_image is now Out[CompressedImage] — anything outside this repo wiring ("color_image", Image) against a go2 graph must retype or decode.
• go2 blueprints referencing the shared perception modules now use the Compressed* subclasses.

How to Test

CI=1 uv run pytest   # full suite green locally (2507 passed)
PYTEST_VERSION=1 uv run python -m dimos.protocol.pubsub.benchmark.tool_replay_bench \
  --blueprint unitree-go2-basic --sink decode --duration 30 --out /tmp/bench

dimos --replay run unitree-go2 shows the camera in rerun via EncodedImage (no pixel decode in the viewer path).

Contributor License Agreement

• I have read and approved the CLA

[codecov]

❌ 1 Tests Failed:

Tests completed	Failed	Passed	Skipped
2847	1	2846	89

View the full list of 1 ❄️ flaky test(s)

dimos.e2e_tests.test_dimsim_spatial_memory::test_go_to_the_bed

Flake rate in main: 17.27% (Passed 115 times, Failed 24 times)

Stack Traces | 525s run time

lcm_spy = <dimos.e2e_tests.lcm_spy.LcmSpy object at 0x7a72994111c0>
start_blueprint = <function start_blueprint.<locals>.set_name_and_start at 0x7a7297deb600>
human_input = <function human_input.<locals>.send_human_input at 0x7a7297deb6a0>
dim_sim = <dimos.e2e_tests.dim_sim_client.DimSimClient object at 0x7a729917bb00>
explore_house = <function explore_house.<locals>.explore at 0x7a7297debba0>

    @pytest.mark.self_hosted_large
    def test_go_to_the_bed(lcm_spy, start_blueprint, human_input, dim_sim, explore_house) -> None:
        start_blueprint(
            "run",
            "unitree-go2-agentic",
            simulator="dimsim",
        )
        lcm_spy.save_topic(".../McpClient/on_system_modules/res")
        lcm_spy.wait_for_saved_topic(".../McpClient/on_system_modules/res", timeout=1200.0)
    
        explore_house()
    
        human_input("go to the bed")
    
>       lcm_spy.wait_until_odom_position(-3.567, -1.332, threshold=2, timeout=180)

dim_sim    = <dimos.e2e_tests.dim_sim_client.DimSimClient object at 0x7a729917bb00>
explore_house = <function explore_house.<locals>.explore at 0x7a7297debba0>
human_input = <function human_input.<locals>.send_human_input at 0x7a7297deb6a0>
lcm_spy    = <dimos.e2e_tests.lcm_spy.LcmSpy object at 0x7a72994111c0>
start_blueprint = <function start_blueprint.<locals>.set_name_and_start at 0x7a7297deb600>

dimos/e2e_tests/test_dimsim_spatial_memory.py:32: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
dimos/e2e_tests/lcm_spy.py:167: in wait_until_odom_position
    self.wait_for_message_result(
        predicate  = <function LcmSpy.wait_until_odom_position.<locals>.predicate at 0x7a7297debd80>
        self       = <dimos.e2e_tests.lcm_spy.LcmSpy object at 0x7a72994111c0>
        threshold  = 2
        timeout    = 180
        x          = -3.567
        y          = -1.332
dimos/e2e_tests/lcm_spy.py:153: in wait_for_message_result
    wait_until(
        event      = <threading.Event at 0x7a72989b0fe0: unset>
        fail_message = 'Failed to get to position x=-3.567, y=-1.332'
        listener   = <function LcmSpy.wait_for_message_result.<locals>.listener at 0x7a7297debc40>
        predicate  = <function LcmSpy.wait_until_odom_position.<locals>.predicate at 0x7a7297debd80>
        self       = <dimos.e2e_tests.lcm_spy.LcmSpy object at 0x7a72994111c0>
        timeout    = 180
        topic      = '/odom#geometry_msgs.PoseStamped'
        type       = <class 'dimos.msgs.geometry_msgs.PoseStamped.PoseStamped'>
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

predicate = <bound method Event.is_set of <threading.Event at 0x7a72989b0fe0: unset>>

    def wait_until(
        predicate: Callable[[], bool],
        *,
        timeout: float,
        interval: float = 0.1,
        message: str | None = None,
    ) -> None:
        """Poll ``predicate`` until it returns truthy or ``timeout`` elapses."""
        deadline = time.monotonic() + timeout
        while time.monotonic() < deadline:
            if predicate():
                return
            time.sleep(interval)
>       raise TimeoutError(message or f"Timed out after {timeout}s waiting for condition")
E       TimeoutError: Failed to get to position x=-3.567, y=-1.332

deadline   = 290068.968953767
interval   = 0.1
message    = 'Failed to get to position x=-3.567, y=-1.332'
predicate  = <bound method Event.is_set of <threading.Event at 0x7a72989b0fe0: unset>>
timeout    = 180

.../utils/testing/waiting.py:35: TimeoutError

To view more test analytics, go to the Test Analytics Dashboard
_{📋 Got 3 mins? Take this short survey to help us improve Test Analytics.}