FIX: Restore airt.cyber E2E + azure-ai-evaluation partner contract

[romanlutz]

Two unrelated CI failures on main, surfaced together because the GitHub check-run annotations only carried wrapper messages. Full investigation lives in the commit body; short version below.

End-to-End scenario tests (AzDO build #11909)

Every test_scenario_with_pyrit_scan[*] parametrization fails with Server not available at http://localhost:8000 because PR #1545 turned pyrit_scan into a thin client of a separate pyrit_backend server, and the e2e test never started one. The airt.cyber attribution in the GitHub annotation is just whichever scenario pytest printed last — every airt scenario, benchmark.adversarial, garak.encoding, etc., fail the same way.

Fixing the server launch surfaced a second latent failure: PR #1785 (scenario technique consolidation) made GET /api/scenarios/catalog/<name> instantiate every scenario class. Cyber.__init__ calls _build_cyber_strategy(), which requires AttackTechniqueRegistry to be populated — but the scenario_technique initializer was not in the test config, so the catalog GET returned 500 before any scenario run could begin.

Fixes:

• tests/end_to_end/conftest.py (new): session-scoped autouse fixture that launches pyrit_backend via ServerLauncher and tears it down on exit. Idempotent if a backend is already healthy.
• tests/end_to_end/test_config.yaml: declare scenario_technique as a backend startup initializer so the catalog endpoint can instantiate scenarios that rely on AttackTechniqueRegistry.

Verified locally end-to-end with a dummy API key: the catalog endpoint succeeds, the scenario runs both strategies (2/2 attacks), and only the final OpenAI request fails with 401 (fake key). On AzDO with real Key Vault credentials the scenario will pass.

Partner integration test (AzDO build #11908)

test_scorer_identifier_importable fails with:

ImportError: cannot import name ''ScorerIdentifier'' from ''pyrit.identifiers''

(The AzDO "exit code 2" is the script wrapper, not a pytest collection error — pytest itself reported 1 failed / 97 passed / 3 skipped.)

PR #1387 collapsed ScorerIdentifier/AttackIdentifier/ConverterIdentifier/TargetIdentifier into ComponentIdentifier with no deprecation alias. azure-ai-evaluation''s _rai_scorer.py still does from pyrit.identifiers import ScorerIdentifier and uses it as a return-type annotation (verified against the live partner source), so the test correctly flagged a real partner contract break.

Fix:

• pyrit/identifiers/__init__.py: PEP 562 __getattr__ returns ComponentIdentifier for the name ScorerIdentifier and emits print_deprecation_message(removed_in="0.16.0") per the project deprecation policy. No partner code change required to keep azure-ai-evaluation working; the alias buys them a normal deprecation window to migrate.

Verification

• tests/partner_integration: 98 passed, 3 skipped (was 97 passed, 1 failed, 3 skipped)
• tests/unit/identifiers: 210 passed
• tests/unit/cli/test_pyrit_scan.py + test_scenario_service: clean
• tests/end_to_end/test_scenarios.py[airt.cyber] locally: backend starts, catalog 200s, scenario runs to LLM call (failed only on fake-key 401)

Out of scope

The 3 test_all_datasets.py HuggingFace/GitHub fetch failures in build #11909 (flaky 3rd-party HTTP) are unrelated to either root cause and are being handled in a separate session.