feat: agentops assert run + agentops redteam run as active CI gates

[placerda]

What

Turns ASSERT (open-source assert-ai framework) and the Foundry/PyRIT AI Red Teaming agent into active, gated CI steps that AgentOps orchestrates end-to-end — instead of just consuming pre-generated artifacts via assert_path: / redteam_path:.

Also rewrites the README Overview to position AgentOps as an open-source orchestration framework + CLI rather than a feature bullet list.

New commands

• agentops assert run — subprocess wrapper around the assert-ai CLI. Parses metrics.json + scores.jsonl, writes a normalized summary at .agentops/assert/latest.json, and exits 2 on any policy violation (unless --no-gate).
• agentops redteam run — invokes azure.ai.evaluation.red_team.RedTeam against an Azure OpenAI deployment, Foundry agent, or HTTP endpoint. Aggregates per-category and per-strategy attack-success-rate into .agentops/redteam/latest.json and exits 2 when ASR exceeds the configured threshold (unless --no-gate).

Both runners produce artifacts that agentops doctor --evidence-pack ingests automatically.

Schema additions (agentops.yaml)

assert:
  cli: assert-ai
  config: assert/eval_config.yaml
  fail_on_violations: true

redteam:
  target:
    type: model_deployment
    deployment: gpt-4o-mini
  risk_categories: [violence, hate_unfairness, sexual, self_harm]
  attack_strategies: [base64, rot13]
  num_objectives: 5
  fail_on_attack_success_rate: 0.2

assert / redteam are YAML aliases for the Python fields assert_run / redteam_run (the keyword assert cannot be a Python identifier).

Docs

• Tutorial: docs/tutorial-prompt-agent-quickstart.md step 10 rewritten as 10a (ASSERT) / 10b (Red Team) / 10c (evidence pack pickup), with install steps for assert-ai and azure-ai-evaluation[redteam].
• README: new tagline, new Overview as a six-step orchestration narrative (Evaluate → Probe → Diagnose → Gate → Prove → Learn-from-production), updated Foundry boundary table with a new "Probe safety" row.

Tests

• New file tests/unit/test_assert_and_redteam_runners.py — 24 tests covering schema aliases, run-output discovery, dimension summarization, totals aggregation, target-callback shapes, normalized JSON writing, gating, and CLI smoke (missing config / missing dependency / explain manuals).
• Full suite: 921 passed, 1 skipped.

Exit-code contract preserved

• 0 — all gates passed
• 2 — threshold, ASSERT violation, or red-team ASR gate failed
• 1 — runtime / configuration error

Scope notes

• Active runners are additive. Existing assert_path: / redteam_path: evidence-only paths still work.
• AgentOps does not reimplement ASSERT or PyRIT — it shells out to / imports the upstream tools and normalizes their output.
• No changes to agentops doctor, cockpit, or eval runner internals beyond the new normalized JSON inputs.

Follow-ups (out of scope)

• Cockpit cards for ASSERT / Red Team latest runs.
• Patch release to PyPI (currently v0.3.13; user tutorial workspace will need v0.3.14 to consume).