feat(sae): SAE steering — delta clamps and full reconstruction

[RhizoNymph]

Brings SAE-based steering into the integration line: per-(layer, hook) SAE feature surgery composing with the existing additive steering, capture, and patching stack.

What this adds

SAE delta tier (sae_steering.py, SAEClampManager): encode the live residual with a registered SAE, replace a chosen set of feature activations with caller-supplied clamp values (absolute or additive), and add the resulting decoder-direction delta back into the residual stream. Clamp tables use row 0 as the no-op sentinel, rows 1/2 as global prefill/decode tiers, rows 3+ per-request; global and per-request clamps validate against overlap in both directions and set_global_clamps is atomic.

SAE full-reconstruction tier (sae_full_reconstruction.py, SAEFullReconstructionManager): replace the residual with decode(activate(encode(h))) + b_dec, optionally clamping the activation before decode (Scaling Monosemanticity / Golden Gate semantics). Rows are gated per-site by an active-row table (a row owned by another module's site never reconstructs here) and keyed by FR-only phase hashes so additive state can't route tokens into a reconstruction row in the wrong phase.

Runtime integration: both tiers ride the existing machinery — SamplingParams.sae_clamp_specs / sae_full_reconstruction_specs with cached phase hashes folded into the steering config hashes (prefix-cache aware), three independent scheduler capacity pools (additive / SAE clamp / full-recon) under the config-pool backpressure model, admission/release/streaming-refresh/preemption-resumption lifecycle in the steering runner mixin, TP/PP via the same collective_rpc discipline, and SAE module registration (manifest + encoder/decoder weights) through the steering modules registry and /v1/steering/modules router with kind: sae_delta | sae_full_reconstruction.

Hook semantics: SAE surgery at the block output operates on the true block output residual + hidden_states (the #198/#202 post_block consolidation applied to SAE), writing back through the residual summand as a delta so untouched tokens stay bit-identical; the sum is only formed when SAE buffers are attached at the site.

Notes for review

• This branch already contains feat/integration (merged; 898 conflicts resolved with integration's framework as canonical), so the diff here is purely the SAE feature.
• SteeringManager.register_config gains optional content-hash row aliasing (identical vector content shares a physical row); implicit aliasing applies to prefill rows only, keeping decode rows 1:1 for the per-row scale/monitor machinery.
• GPU-dependent suites (tests/models/language/generation/test_sae_*_real_weights.py, test_steering.py, tests/v1/core/test_steering_config_reservation.py) still need a pass on GPU hardware.