Integrate upstream/main into feat/integration

[RhizoNymph]

What

Merges ~846 commits from upstream/main (through a331589394, 2026-06-18) into the integration fork.

Why

Keeps the steering/capture fork current with upstream. Merge-base was ~3.5 weeks stale (2026-05-24).

Conflict resolution

81 files overlapped; 15 conflicted. Notable resolutions:

• LLM internals moved upstream: LLM private methods were extracted into a new OfflineInferenceMixin (vllm/entrypoints/offline_utils.py). Ported the inline-steering hook (_maybe_pack_inline_steering + its _add_request call site) into the mixin; removed the now-duplicate block from llm.py (constructor capture-consumer logic auto-merged and is intact).
• Scheduler: kept _set_request_block_hash_steering_overrides; took upstream's new schedule(throttle_prefills=...) + current_step increment.
• GGUF removed from tree: upstream extracted GGUF into the external vllm-gguf-plugin. Accepted the deletions; repointed the extra-quant optional dependency at the RhizoNymph/vllm-gguf-plugin fork (adds gemma4 support).
• Removed models: accepted upstream removal of dots1 and internlm2_ve (our changes there were only hooks).
• Remaining conflicts were additive (kept both sides).

Runtime-validated on node0 (RTX 3090): engine initializes and generates correctly (Qwen3-4B). The ~67 textually auto-merged files (incl. gpu_model_runner.py) build and run but warrant deeper capture/steering validation.

[RhizoNymph]

v2 model runner steering + capture merged in + GPU re-validated

Merged `feat/integration` (carrying the v2-model-runner steering/capture port, PR #185) into this branch — merge commit `4649466aa7`.

Conflict resolved in one file, `vllm/v1/worker/gpu/model_runner.py` (3 regions, all "upstream-added-X coincides with v2-port-added-Y on the same lines"):

1. Imports — kept upstream's `set_default_max_concurrency` and the v2 `CaptureRunnerMixin` import.
2. `dispatch_cg_and_sync_dp(...)` — kept both `need_eager=... or capture_pending` and upstream's new `num_active_loras=num_active_loras`.
3. `ModelRunnerOutput(...)` ctor — upstream moved `kv_connector_output` to a post-build assignment (~L1534), so dropped the duplicate ctor arg (would reference an undefined local) and kept only `capture_results=self._drain_capture_results()`.

No further accommodation needed: all capture/steering call-sites and both `init*_state` hooks survive the merge, and the mixin imports resolve against upstream's utils reorg.

Validation (Qwen3-0.6B, dummy weights, `enable_steering` + filesystem capture consumer):

• 19/19 v2 capture+steering CPU glue tests pass.
• GPU, all PASS: single-node eager, single-node cudagraph (FULL+PIECEWISE), TP=2 and PP=2 cross-node (Ray + NCCL over bond0).
• Steering now works on the V2 runner with no `VLLM_USE_V2_MODEL_RUNNER=0` workaround (Qwen3 on its default V2 path): output shifts under a vector and clears to baseline; TP reports both ranks, PP reports correct per-stage layer ownership; capture tensors valid and written to the owning rank/stage.
• Capture-registry resilience fix confirmed working (broken third-party consumer entry points are skipped with a warning instead of crashing admission).

Non-blocking: a pre-existing `capture_residual` custom-op aliasing deprecation warning (PyTorch 2.12) in the capture data plane — not introduced by this merge.

AI assistance was used for this work.