Forward-merge release/0.1 into main - #125

#### Overview

Implements streaming backed by NVIDIA NeMo Relay for NVIDIA NeMo Fabric
runtimes, promoting the streaming proof of concept from #104 into the supported
Python SDK surface.

`Runtime.invoke_stream()` exposes raw ATOF records while an invocation is running
and keeps the normalized terminal `RunResult` out of band. Streaming is **opt-in**:
callers pass `streaming=True` to `start_runtime(...)` (with NeMo Relay enabled) to
provision it, so NeMo Relay-enabled runtimes that do not stream pay no additional cost
and keep their existing telemetry shape. The streaming API change is additive — a
new optional `streaming` parameter — and introduces no new dependencies.

#### Details

- Add an SDK-owned loopback HTTP listener for NeMo Relay NDJSON/ATOF delivery with a
  bounded, byte-budgeted record queue (16 MiB total / 1 MiB per record), TCP
  backpressure, and a bounded post-turn drain window.
- Gate provisioning behind `start_runtime(..., streaming=True)`: a NeMo Relay-enabled
  runtime binds the loopback listener and injects the `nemo-fabric-stream` ATOF
  sink only when streaming is requested. Requesting `streaming=True` without NeMo Relay
  enabled raises `FabricConfigError`.
- Add `Runtime.invoke_stream()`, `Runtime.supports_streaming`,
  `InvokeStream.result()`, and explicit early-exit cleanup via
  `InvokeStream.aclose()`.
- Configure the listener host through an environment variable so deployments can
  override the loopback default when needed.
- Correlate records to a single invocation by ATOF scope tree: Deep Agents and
  Hermes stamp a request ID on the turn's root scope; gateway harnesses match on
  NeMo Relay's turn role and index. The Hermes correlation scope is added only for
  streaming runs, so non-streaming NeMo Relay runs keep their prior trajectory shape.
- Surface otherwise-silent failures as `RuntimeWarning`s: no NeMo Relay connection
  reached the listener, data arrived but no record matched the active turn, or a
  chunked upload was interrupted mid-turn. A clean stream completion does not warn.
- Validate streaming requests before returning an `InvokeStream`, make stream
  iteration cancellation-safe, and allow a manual `Runtime.stop()` to finalize a
  completed stream after `result()`.
- Document the public streaming contract, terminal failure semantics, adapter
  behavior, runnable examples, and the generated Python API reference.

#### Validation

- Focused streaming, adapter, generated-reference, and Fern-sync tests: 135 passed.
- `uv run --no-sync pre-commit run --all-files --show-diff-on-failure` — all hooks
  passed, including Cargo/uv lockfile freshness, attribution generation, and the
  dependency license diff.
- `just docs` — passed; Fern reported only the expected unauthenticated redirect
  check warning.
- `just test-python` — 455 passed and 44 skipped locally. Two failures and two
  setup errors remain in `test_installed_adapter_discovery.py` because uv's local
  macOS Python 3.13 virtual environments cannot load `libpython3.13.dylib`; the
  streaming-focused tests pass independently.
- End-to-end streaming turns run under
  `warnings.simplefilter("error", RuntimeWarning)` and assert non-empty records:
  - Deep Agents — `tests/e2e/test_deepagents.py` (persistent host, mock model).
  - Claude and Codex gateways — live tests in `tests/e2e/test_claude.py` and
    `tests/e2e/test_codex.py`, validating gateway turn correlation.

#### Where should the reviewer start?

Start with `python/src/nemo_fabric/streaming.py` for the listener, turn
correlation, and stream lifecycle; then `python/src/nemo_fabric/client.py` for the
`streaming=True` opt-in gate and `python/src/nemo_fabric/runtime.py` for eager
validation and manual shutdown. The focused behavioral coverage is in
`tests/python/test_streaming.py`, with end-to-end coverage under `tests/e2e/`.

#### Related Issues: (use one of the action keywords Closes / Fixes / Resolves / Relates to)

- Supersedes the implementation POC in #104

- [x] I confirm this contribution is my own work, or I have the right to submit it under this project's license.
- [x] I searched existing issues and open pull requests, and this does not duplicate existing work.



## Summary by CodeRabbit

* **New Features**
  * Added NeMo Relay-backed streaming to the Python SDK via `Runtime.invoke_stream()` with an `InvokeStream` async iterator of raw ATOF records and a terminal `await stream.result()`.
  * Added `Runtime.supports_streaming` and streaming-aware runtime startup/stop lifecycle handling.
  * Introduced typed v2 NeMo Relay ATOF sink configuration via file/stream sink models.

* **Documentation**
  * Added the Streaming API reference and updated Python and integration guides with end-to-end streaming workflows.
  * Updated NeMo Relay ATOF examples to use the typed v2 `sinks` model.

* **Tests**
  * Expanded/updated streaming, documentation contract, and adapter configuration tests to match the new streaming and typed sink behavior.

Authors:
  - Yuchen Zhang (https://github.com/yczhang-nv)

Approvers:
  - Anuradha Karuppiah (https://github.com/AnuradhaKaruppiah)
  - Ajay Thorve (https://github.com/AjayThorve)

URL: #109