docs(http-agent): deduplicate judge env var block (#129)

[Dongbumlee]

Closes #129.

Summary

Doc-only validation pass for docs/tutorial-http-agent.md against current develop (374 lines). One drift fixed.

Drift fixed

Issue	Evidence
Prerequisites set both AZURE_OPENAI_DEPLOYMENT and AZURE_AI_MODEL_DEPLOYMENT_NAME to the same value.	_model_config() reads them as fallbacks of each other — setting both is redundant. Same fix shipped in PRs #158 / #159 for the Copilot-skills and agent-workflow tutorials.

Runtime verification (hybrid testing, Option C)

The ACA deploy in Section 2 is heavy infra. Used a two-part approach:

Part 1 — Azure CLI flag check (covers Section 2)

az containerapp up on current containerapp extension v1.3.0b4 still accepts every documented flag:

Flag	Present in --help
--name -n	✅
--resource-group -g	✅
--location -l	✅
--source	✅
--target-port	✅
--ingress (external/internal)	✅

Part 2 — Local FastAPI stand-in (covers Sections 1, 3–6 end-to-end)

Wrote the tutorial's exact app.py and ran uvicorn app:app --host 127.0.0.1 --port 8000. All three documented /chat paths returned the documented JSON shape:

// lookup
{"text":"Order ORD-12345 is in transit and expected to arrive tomorrow.","tool_calls":[{"type":"tool_call","tool_call_id":"lookup_1","name":"lookup_order","arguments":{"order_id":"ORD-12345"}}]}
// refund
{"text":"I started a refund for ORD-77821 because it arrived broken.","tool_calls":[{"type":"tool_call","tool_call_id":"refund_1","name":"refund_order","arguments":{"order_id":"ORD-77821","reason":"arrived broken"}}]}
// greeting
{"text":"Hello! I can help with order status, refunds, or connecting you to support.","tool_calls":[]}

agentops eval run against http://127.0.0.1:8000/chat produced all six documented metrics:

| coherence           | 4.000 | ✅ |
| fluency             | 3.000 | ✅ |
| tool_call_accuracy  | 5.000 | ✅ |
| intent_resolution   | 4.000 | ✅ |
| task_adherence      | 0.333 | ❌ |  (real signal — synthetic agent doesn't really adhere)
| avg_latency_seconds | 0.003 | ✅ |

Tool calls were correctly extracted from the HTTP response via tool_calls_field: tool_calls. The greeting row (no tools) correctly showed tool_call_accuracy=n/a. results.json + report.md written under the documented timestamped folder + latest/ mirror.

Out of scope

ACA provisioning itself wasn't exercised. AgentOps doesn't change behaviour based on the URL pointing at ACA vs localhost — only that the URL is reachable and returns the documented JSON shape. Anyone who wants the full deploy verified should run Section 2 verbatim; the doc has no claims about ACA-specific behaviour that the local run cannot exercise.

Tests

Full suite: 346 passed, 1 skipped (with the pre-existing test_cli_platform_invalid_value_fails deselected — Click 8.2 stderr issue on develop, unrelated).

Note for reviewers

Branched directly off current develop. No code changes, no dependencies on other PRs.