# KIP-0015: Multi-Stream Adversarial Testing ## Status Accepted for Milestone 9. ## Summary Milestone 8 added lab-only multi-stream semantics. Milestone 9 adds adversarial multi-stream scenarios, stream-level feature extraction, collapse scanning, stream mutants, audit gates, and generated-backend parity checks. This remains strictly local and lab-only. It does not add SOCKS, VPN mode, HTTP carriers, TLS mimicry, CDN behavior, deployment scripts, production key exchange, external targets, mobile apps, live-network testing, or operational evasion guidance. ## Why Happy-Path Echo Is Insufficient A simple multi-stream echo can pass even if all profiles expose the same stream behavior. Adversarial stream scenarios force the scheduler, flow-control model, close/reset lifecycle, and trace metadata to show whether generated profiles still differ under pressure. The goal is not to prove real-world resistance. The goal is to catch local regressions where stream polymorphism collapses into fixed observable patterns. ## Scenario Definitions - `balanced_interleave`: opens four streams, sends similar chunk sizes, interleaves progress fairly, and closes all streams normally. - `bulk_vs_interactive`: opens one large bulk stream plus small interactive streams and checks that `interactive_first` profiles show interactive progress before bulk work. - `blocked_stream`: exhausts one stream window, verifies that only that stream is blocked, and verifies that other streams continue. - `session_window_exhaustion`: exhausts the session window, verifies global backpressure, and verifies recovery when a window update is represented. - `reset_midstream`: resets one stream after partial data and verifies that other streams continue. - `close_race`: closes one stream while another is active and verifies the lifecycle remains valid. - `uneven_stream_sizes`: mixes tiny, medium, and large streams and verifies that small streams are not starved in the lab scheduler model. ## Stream-Level Features The stream adversary extracts payload-free features: - stream count - stream open order - stream close order - reset count - reset position bucket - window-update count - backpressure event count - blocked stream ratio - session-blocked count - interleaving score - fairness score - largest-stream dominance ratio - scheduler decision pattern - priority-progress ratio - stream ID pattern bucket - close/reset outcome pattern Traces do not include payload contents, raw frames, secrets, proofs, keys, or external target data. ## Collapse Scanner The collapse scanner checks whether too many profiles share the same observable stream behavior. It flags suspicious stability for: - same stream ID sequence or encoding bucket - same window-update rhythm - same scheduler decision pattern - same close/reset outcome - the same composite stream behavior across profiles Some same-shape behavior is expected for a specific scenario, such as balanced open/close order. The scanner treats those as context rather than automatic failures. ## Stream Mutants The test-only stream mutants simulate regressions: - `fixed_stream_id_strategy`: all profiles share one stream ID strategy or encoding. - `fixed_window_update_policy`: all profiles share window-update thresholds or rhythm. - `fifo_scheduler_only`: profiles collapse to FIFO behavior regardless of configured priority policy. - `fixed_reset_close_policy`: all profiles share reset and close behavior. - `no_backpressure`: runtime behavior ignores flow-control backpressure. - `padding_only_stream_diversity`: stream behavior is fixed and only padding varies. The audit gate passes only when these mutants are detected as regressions. ## Audit Gates Root `kcheck --quick` now includes: - `multi_stream_adversarial_scenarios` - `multi_stream_collapse_resistance` - `multi_stream_mutant_detection` Generated-backend audit includes: - `multi_stream_generated_backend_parity` The generated-backend gate relies on generated module tests plus generated trace metadata. Generated modules run adversarial scenario tests against their profile-specific static profile. ## Commands ```bash go run ./cmd/kcheck --quick go run ./cmd/kcheck streamadversary --quick go run ./cmd/kcheck streamadversary --full --out testdata/audit/stream-adversary.json go run ./cmd/kcheck codegen --quick ``` ## Limitations - This is a deterministic lab model, not a production transport. - Scenarios are local and synthetic. They do not model real network loss, delay, middleboxes, or live censorship systems. - Feature extraction is heuristic and payload-free. It can detect known collapse modes but cannot prove undetectability. - Generated code still reuses shared lab helpers for safe IO, stream session logic, trace output, and test harnesses. - Passing the stream adversary audit does not imply production readiness, real-world robustness, or censorship resistance. ## Next Layer Milestone 10 extends this adversarial model to proxy-style relay intents, synthetic targets, target-induced backpressure, and target error/reset isolation. See [KIP-0016: Lab Proxy Semantics](KIP-0016-lab-proxy-semantics.md).