[LeeO86]

This is M4 of IMPLEMENTATION_PLAN.md and the preparation for 1.1.0 (M3 and M4 ship together). It is based on #6: merge #6 first, then this PR, which then targets main. Release notes for the platform: docs/release-notes-1.1.0.md.

Commits

1. AMWA nmos-testing in CI: IS-04-01, IS-05-01, IS-05-02 and BCP-007-03-01 with an nmos-cpp registry (DNS-SD off), adapted from mxl-audio-mixer. tests/nmos/amwa-expected-off.json gives the reason for each of the 99 non-Pass results:
   • set-up and tear-down steps;
   • IS-10 authorization;
   • the RTP-only parts of IS-05;
   • optional BCP-004/002 details;
   • two manual checks.
2. Fix, found by AMWA: the receiver's IS-04 caps lost media_types (BCP-007-03-01 test_07 failed).
3. Monitoring and demo:
   • Standard process metrics on /metrics.
   • Grafana dashboard deploy/grafana/mxl-scopes.json with provisioning, plus a Prometheus config.
   • Compose demo docker/docker-compose.demo.yaml, checked with docker compose config only, because its registry announces itself over mDNS.
4. deploy/k8s/deployment.yaml: requests 500m CPU and 256Mi; limits 2 CPUs and 768Mi. docs/cuda-design.md.
5. 1.1.0 itself:
   • version 1.1.0 (the Container workflow checks the tag against it);
   • release notes;
   • docs/performance.md "M4": the platform's numbers, memory per mode, the soak, AMWA;
   • spec TODOs resolved (ports confirmed, measured memory and CPU);
   • plan, README, CHANGELOG, AGENTS.md.

Numbers

Platform (i9, 1.0.0, 25 Hz, no preview; measured by the platform):

Mode	CPU (core)	RSS
luma	0.10	63–116 MB
ycbcr-parade	0.16	87 MB
rgb-parade	0.19	103 MB
rgb-overlay	0.26	114 MB
ycbcr-overlay	0.18	114 MB
luma with a key	0.17	—

The keyed scope over a picture passed on test-all with multiviewer 1.4.6 (CUDA): compose 1.56 ms, 0 late and 0 missed. This closes the M2 exit check.

Lab memory (1.1.0): 80–160 MB of the scope's own (anonymous) memory per mode and configuration. The shared MXL grains it maps come on top, and the built-in MediaMTX takes about 41 MB.

Soak, 1 h, two instances on one node (luma 25 Hz; RGB parade 25 Hz with a key; both with the preview):

• RSS stays flat; 0 torn reads.
• Instance A: 1 late grain, at start-up.
• Instance B: 16 late and 1 missed, all at start-up or in two bursts while other work ran on the host; none in the last 43 minutes.

Review fixes (e0c4a00, c45182b; rebased on feat/m3 51eeaaa)

• AMWA reasons: the IS-04-01 DNS-SD tests are disabled by ENABLE_DNS_SD = False, not IS-10; IS-05-01 test_01–08/33 are "replaced by an auto test". docs/performance.md and the release notes say AMWA does not cover IS-04 registration and heartbeats (only the stub registry in scope.sh does). No unicast DNS-SD added.
• process_cpu_seconds_total uses the number() helper (unit test: no exponent). Docs: RSS and process_resident_memory_bytes exclude the MediaMTX child.
• k8s: request 384Mi; a note that the 768Mi limit is for 1080p and UHD with a key needs about 1Gi.
• Soak re-run (1 h, after the 1.1.0 part 1 (M3): WebRTC preview, Vue UI, scope widget, Blend2D texts #6 fixes, pinned, host otherwise only pinned work): A luma 0 late / 0 missed, 0.41 core, RSS 341 → 357 MB; B RGB parade + key 1 late (start-up) / 0 missed, 0.65 core, RSS 514 → 515 MB; 0 skipped, 0 torn.

Open

• Desktop-class numbers with the preview: the platform measures them after rollout.
• The M3 exit check on a real operator screen.
• An incident during this work, below.

Incident. While trying a local AMWA run (2026-10-23, about 06:02–06:06Z), I started an nmos-cpp registry on the lab host's network for about 2.5 minutes. Its mDNS responder announced it, and at least one other node on the network (id de10381b-…) registered to it. I stopped it. AMWA now runs only in CI, and AGENTS.md warns against running a registry on a shared lab network.

🤖 Generated with Claude Code

https://claude.ai/code/session_01St7QwmonpoG9EEEqaLFXLi