feat(rtmg): spectral steering (on models/backend)

[ryanontheinside]

feat(rtmg): spectral steering, re-expressed in the registry-driven control surface

Merge order

Stacked: opened against ryanontheinside/feat/models/spectral/spectral-control. Merge #224 -> #220 -> #229 -> spectral-control before this.
Stack: main <- #224 <- #220 <- #229 (models/backend) <- spectral-control <- this.
Supersedes #163 (same feature, ported from the pre-seam ryanontheinside/feat/spectral/demo-on-spectral-control; close that one in favor of this).
Heads-up for whoever merges this after the mrt2 (#230) / sa3 (#231) leaves (or vice versa): acestep/streaming/families.py will conflict — this PR extends the ACE factory + knob universe, those add their families to the same registries. Resolution is mechanical (keep both).

The demo/control surface for the spectral-control backend underneath: four perceptual steering knobs (bright / warm / rough / density) plus runtime-allocatable manual slots that fire any vector in the probe catalog at an operator-chosen (layer, step, alpha).

What's here (two commits)

4c3eff1 tab rename. "Channels" tab -> "Experimental"; section labels become "channel highlights" / "channel groups".

dd556b8 spectral steering. The original feature, re-expressed in the post-refactor architecture instead of the old hand-rolled knob banks:

• Knob registry (acestep/streaming/knobs.py): pure steering_axis_spec() / manual_slot_specs() factories (same pattern as lora_strength_spec). ACEStepBackend.knob_specs() appends them when the session's checkpoint has a vector bundle, so /api/knobs-shaped manifests, the wire ready frame, MCP list_knobs, and DynamicKnobPanel all pick them up with zero per-transport code.
• Backend (ace_backend.py): owns the SteeringController (constructed in families._make_acestep via ensure_steering_vectors(checkpoint)); _sync_steering() pushes configs on snapshot change (pipeline identity is part of the key so a steps_override rebuild re-arms), sitting next to _sync_channel_guidance in _prepare_tick. New Capabilities.steering bit, true only when vectors loaded.
• Session (session.py): manual_slot_add / manual_slot_pop ops gated by @requires_capability("steering", ...); KnobState + the cached spec map mirror the controller's LIFO slots; ManualSlotCount bus event published on success AND refusal; steering_payload() (count / cap / available) shared by snapshot and the ready frame.
• Wire contract (protocol.py): manual_slot_add / manual_slot_pop commands (requires="steering"), manual_slot_count event, three wire-optional ready fields. wireContract.gen.ts regenerated; all drift guards green.
• MCP (mcp_server.py): add_manual_slot / pop_manual_slot / list_manual_steering_vectors tools; list_knobs and knob validation now prefer the session snapshot's backend-owned knob_manifest (the only projection that can carry backend-specific knobs), falling back to the static registry for older servers.
• Client: SDK stays store-free per Backend-owned control contract + demon-client SDK #220's convention — manualSlotCount / manualSlotCap / steeringAvailable are RemoteBackend instance fields + a typed manual_slot_count switch case; the host mirrors them in useStartSession. Senders typed against the generated command types. VoiceTile grows the steering knob rack + manual-slot rows with +/- controls; tooltips in SliderTile; SLIDER_META pre-registers man_*_{1..16}.

Deliberately dropped from the original commit: the unrelated config.json default LoRA/prompt swaps (funkv1 / Phonk, prompt A) — debugging defaults that rode into #163.

Validation

• 149/149 unit tests (incl. wire-contract drift guards over the new commands/event/capability and the knob-homonym guard over the extended universe).
• Web: tsc clean, next build clean, vitest unit 35/35, replay tier 6/6 (bit-exact).
• Offline smoke against the real cached v15-turbo bundle: 144-cell catalog loads, auto-axis + manual configs build correctly (sign/layer-offset/fractional-step semantics intact).
• Golden harness: pending a free GPU window; will be run at this tip and reported here.