fix(ai_agents): bump example graphs' pinned model to gpt-realtime-2.1

[fra-shipper]

Summary

voice-assistant-realtime, voice-assistant-companion and demo each pin "model": "gpt-realtime" explicitly in their tenapp/property.json, so they don't pick up whatever default openai_mllm_python ships internally — they always use exactly the string written in the graph. OpenAI announced gpt-realtime's deprecation on 2026-07-20 with shutdown on 2027-01-20; the GA replacement is gpt-realtime-2.1 (per developers.openai.com/api/docs/deprecations). This bumps the three pins to match, and updates the two README.md snippets that document the same property.json fields so the docs stay in sync.

demo/tenapp/property.json also pins "gpt-realtime-mini" on a separate node (a different agent config in the same file). That model's correct replacement is gpt-realtime-2.1-mini, not gpt-realtime-2.1, so it is intentionally left untouched here.

demo's own web UI (frontend/src/app/api/agents/start/graph.ts) sends its own OPENAI_REALTIME_MODEL constant as properties.v2v.model on every "OpenAI GPT Realtime" (va_openai_v2v) launch, and the backend's mergeProperties path overwrites the property.json node value with whatever the frontend sends. That constant also pinned "gpt-realtime", which would have silently defeated the property.json bump for anyone launching the demo through its own UI, so it is bumped here too. The "OpenAI GPT Realtime 1.5" (va_openai_v2v_1_5) and "OpenAI GPT Realtime Mini" (va_openai_v2v_mini) dropdown options are separate, explicitly versioned UI choices and are left pointing at gpt-realtime-1.5 / gpt-realtime-mini respectively — same distinction already made for the mini node above. I checked developers.openai.com/api/docs/deprecations directly: gpt-realtime-1.5 is not itself deprecated (it only appears as the replacement target for older 2025 preview snapshots), so collapsing it into 2.1 would silently break that labeled choice.

Context

#2261 migrates openai_mllm_python's realtime client to the GA wire protocol and, as part of that, bumps the extension's own internal config default to gpt-realtime-2.1. Its "Not covered" section flags that these three shipped example graphs override that default explicitly and calls it "worth a separate pass." This PR is that pass, scoped to the model-pin freshness issue across both the property.json graphs and the demo frontend that overrides them. Refs #2261 — it does not itself migrate the wire protocol.

Scope note (please read before reviewing)

This is a freshness/consistency fix only, not a functional fix. As of main, openai_mllm_python's realtime/connection.py still sends the header OpenAI-Beta: realtime=v1 to /v1/realtime. OpenAI retired the Realtime Beta API on 2026-05-12 and now rejects that combination outright ("The Realtime Beta API is no longer supported. Please use /v1/realtime for the GA API."), per #2261's own live verification against api.openai.com. So none of these three example graphs' OpenAI realtime path currently connects, independent of which model string is pinned — this change has no functional effect on main today. Its value is that once the GA client migration lands (#2261 or equivalent), these examples and the demo UI won't still be pointing at a model that's already in its deprecation window.

Testing

This is a static config/constant value bump (JSON string values and one TS const), no branching logic touched, so no unit test applies — there is no existing test harness anywhere in the repo (no jest/vitest config, no *.test.*/*.spec.* files) that exercises these example property.json files' model field or the demo frontend's graph.ts constants. Verified instead by direct inspection, re-run against the current branch state:

• python3 -c "import json; json.load(open(f))" on all three touched property.json files — OK: <path> for all three, confirming the JSON stays well-formed.
• jq empty <path> on the same three files — jq OK: <path> for all three, independent parse confirmation.
• grep -n gpt-realtime ai_agents/agents/examples/demo/tenapp/property.json post-edit — 1533: "model": "gpt-realtime-2.1", and 1691: "model": "gpt-realtime-mini",, confirming exactly the intended node changed and the mini node was correctly left untouched.
• grep -n OPENAI_REALTIME ai_agents/agents/examples/demo/frontend/src/app/api/agents/start/graph.ts — confirms OPENAI_REALTIME_MODEL = "gpt-realtime-2.1" while OPENAI_REALTIME_15_MODEL and OPENAI_REALTIME_MINI_MODEL are unchanged.
• git diff origin/main..HEAD --stat — 6 files changed, 6 insertions(+), 6 deletions(-), confirming a surgical diff with no unintended changes.

Did not attempt a full ai_agents build/run (Taskfile-driven Docker build, requires a live OPENAI_API_KEY to actually connect) — out of scope for a config-value change, and also moot: per PR #2261's own live verification, openai_mllm_python still sends the retired Realtime Beta header, so no real call would succeed on main today regardless of which model string is pinned.