feat(sae-steering): SAE full reconstruction runtime (phase 4 merge)

[RhizoNymph]

Summary

Brings the previously-shipped feat/sae-steering-phase-4-stage-{2,3,3b,4,5} branches into the delta-path mainline so the full-reconstruction (Anthropic Scaling Monosemanticity / Golden Gate Claude) variant runs end-to-end inside the worker. Stages were authored independently and never landed because the delta-path plumbing on feat/sae-steering had moved underneath them; this PR resolves the conflicts against the current delta runtime.

• Stage 2 — per-(layer, hook) full encoder / decoder buffers, per-row clamp tables, shared sae_recon_index, and apply_layer_sae_full_reconstruction shim. Registered as torch.ops.vllm.apply_sae_full_reconstruction so torch.compile treats it as an opaque splitting point.
• Stage 3 / 3b — SAEFullReconstructionSpec request type with hash plumbing folded into hash_steering_config (distinct domain separator from the delta block), SAEFullReconstructionManager parallel to SAEClampManager, worker-mixin lifecycle (admission, prefill→decode transition, release, atomic register-and-attach with rollback).
• Stage 4 — compaction-based CUDA path that runs the encoder / clamp / decoder math only on tokens whose recon_mask is True. cuBLAS-backed matmul on the dense subset; warmup wired into buffer attach so the first-call JIT happens outside any captured forward.
• Stage 5 — Gemma Scope real-weights e2e harness exercising both pure-reconstruction-error-only and clamped-feature paths against a single SAE site of Gemma 2-2B (CUDA + HF token gated).

apply_layer_steering now composes additive → SAE delta → SAE full reconstruction in that order, each behind an independent static hasattr check so the disabled-mode forward still emits zero steering kernels.

Sibling PR

A follow-up adds a global SAE clamp tier as an overlay on the delta path's row-0 sentinel. That PR is independent of this one but expects this PR to land first for the docs cross-references to make sense.

Files

20 files changed, 5666 insertions, 149 deletions. New runtime modules:

• vllm/model_executor/layers/sae_full_reconstruction.py — public op, custom-op registration, layer-hook shim, per-(layer, hook) buffer attach/detach.
• vllm/model_executor/layers/sae_full_reconstruction_kernel.py — compaction-based CUDA path + warmup helper.
• vllm/v1/worker/sae_full_reconstruction_manager.py — per-request row allocator.

Test plan

• pytest tests/v1/test_sae_full_reconstruction_types.py
• pytest tests/v1/worker/test_sae_full_reconstruction_manager.py
• pytest tests/v1/worker/test_sae_full_recon_mixin_integration.py
• pytest tests/model_executor/layers/test_sae_full_reconstruction_*
• pytest tests/models/language/generation/test_sae_full_reconstruction_real_weights.py (CUDA + HF gated)
• Full SAE suite (delta + full-recon + populator + manager + mixin) — 458 passing.