fix(worker): add Dynamo session control for replay sessions

[weireweire]

Summary

• add opt-in Dynamo conversation-aware routing metadata via nvext.session_control
• use each request credit's X-Correlation-ID as nvext.session_control.session_id
• fix session bind lifecycle handling for warmup, profiling continuation, recycled sessions, and branch orchestration
• add debug sampling fields for Dynamo session/request inspection
• fix FIRST_TURN cache-bust marker injection for mid-turn agentic replay resumes by marking seeded turn 0 history exactly once

Validation

• PYTHONPATH=tests uv run --with pytest-asyncio pytest tests/unit/common/config/test_endpoint_config.py tests/unit/common/models/test_endpoint_info.py tests/unit/credit/test_issuer.py tests/unit/credit/test_structs.py tests/unit/timing/strategies/test_agentic_replay.py tests/unit/timing/test_branch_orchestrator_pre_session.py tests/unit/workers/test_worker.py tests/unit/workers/test_worker_cache_bust_injection.py tests/unit/workers/test_worker_cache_bust_adversarial.py -q (186 passed)
• uv run --with 'ruff>=0.14.0,<0.15.0' ruff check src/aiperf/workers/worker.py tests/unit/workers/test_worker.py\n

---

Note

Medium Risk
When routing is enabled, every outbound request body gains merged nvext metadata; cache-bust semantics changed for FIRST_TURN and repeated injection, which can alter KV-cache behavior in benchmarks.

Overview
Adds opt-in Dynamo conversation-aware routing via endpoint flags that emit nvext.session_control on OpenAI-compatible request bodies, with configurable session timeout (default 300s). Config flows from CLI/EndpointConfig into runtime EndpointInfo.

When enabled, the worker merges session_control into the last turn’s extra_body, raw_payload, or decoded payload_bytes, using credit.x_correlation_id as session_id and sending action: bind only on the first request per worker for that session (tracked in _dynamo_bound_session_ids).

Cache-bust fixes: marker injection is idempotent (skips if content already contains the marker), and FIRST_TURN_* targets apply to the seeded first user turn even when credits resume at turn_index > 0 (agentic replay), without duplicating markers on repeated _apply_cache_bust calls.

^{Reviewed by Cursor Bugbot for commit 8c9067d. Bugbot is set up for automated code reviews on this repo. Configure here.}

[github-actions]

Try out this PR

Quick install:

pip install --upgrade --force-reinstall git+https://github.com/ai-dynamo/aiperf.git@8c9067d0bd8d13161606b2ab59307f896b82ae99

Recommended with virtual environment (using uv):

uv venv --python 3.12 && source .venv/bin/activate
uv pip install --upgrade --force-reinstall git+https://github.com/ai-dynamo/aiperf.git@8c9067d0bd8d13161606b2ab59307f896b82ae99

---

Last updated for commit: 8c9067d • Browse code

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 948925a. Configure here.}