feat(C3): add parallel deployment cohort isolation control (3.3.5)

[RicoKomenda]

Summary

Adds 3.3.5 to C3.3 (Controlled Deployment & Rollback) to address a gap where no control prevents cross-contamination between model versions running simultaneously in A/B, canary, or shadow deployments.

New control:

Verify that model versions running in parallel (e.g., A/B tests, canary deployments, shadow deployments) are isolated such that prompt caches, session state, and retrieval context are not shared across cohorts, preventing cross-version data contamination and behavioral inference across deployment groups.

Level: 2

Why this is needed

Canary and A/B deployments route requests to different model versions simultaneously. If shared infrastructure (KV caches, session stores, vector retrieval indices) leaks between cohorts, a user assigned to the stable model can receive responses influenced by the canary model's cached state, and vice versa. This creates both a security boundary violation (canary model state visible to stable model users) and a data integrity problem (behavioral measurements are contaminated across groups).

Existing C3.3 controls cover rollout mechanics (3.3.1) and rollback (3.3.2) but say nothing about isolation between concurrently running versions. This is AI-specific because prompt caches, embedding caches, and retrieval indices are AI-layer constructs with no direct equivalent in standard blue/green web deployments.

Changes

• 1.0/en/0x10-C03-Model-Lifecycle-Management.md: add 3.3.5, fix MD060 separator rows
• 1.0/en/0x93-Appendix-D_AI_Security_Controls_Inventory.md: add entry to AD.13