Add opt-in native Crisis report advisory audit

[Pigbibi]

Audit one frozen Crisis shadow report through an explicit, disabled-by-default Dot entry and write a separate advisory. The existing Crisis strategy calculations, risk/allocation/order controls and artifact writers stay unchanged; the retired Crisis gateway path fails closed, while other scenarios retain their existing behavior.

The caller reuses the pinned Quant ScenarioTaskSpec/TaskClient helpers, persists an exclusive private intent before one submission, and verifies the original task against the frozen report, service resource/audience and request identity. Missing evidence requires a meaningful data gap and human review. Incompatible report identities and extra strategy/authority output fields are rejected. Unknown submission outcomes retain the intent and never retry or fall back.

Validation: 621 offline tests and 117 subtests passed, including the original 576 tests plus 45 review regressions. A1–A3 were independently re-reviewed and closed. Local build and diff checks passed. No real task, model, subscription or production action was exercised.

Scope: exactly six product files. native-audit is optional and pins QuantAIService f318d0ad1183fcf07574f6e00baf8809c8552b8d; the original ai pin and existing dependency records remain unchanged. Eight packages are added to the optional dependency closure, with no existing package upgrades/removals.

Natural CI for head add2efb3dadc9020a15624553207637648df7f18 succeeded: CI run 39277762561, pull_request event, attempt 1. Both actionlint and test jobs passed. The test job reported 507 passed, 114 skipped, 117 subtests passed; pip check, Ruff 0.17.0 and the package build passed.

The 114 native cases, including the 45 review regressions, were collected but skipped: this workflow installs only .[test,ai] and the native fixture uses pytest.importorskip("ai_service.scenarios"). This CI result is not a native-test PASS. The full 621-test native-enabled result above remains separate offline evidence.

Draft adoption limits remain: an explicitly approved optional native CI channel, clean full locked native installation, actual host/grant/profile-to-Dot/worker bindings and authoritative unknown-outcome recovery. The pre-existing Ruff 0.17.0 declaration versus 0.16.6 lock mismatch is unchanged; this pip-based CI does not validate that lock consistency. No dependencies or additional workflow files were changed to hide these limits.