Feature abot dit

[DmitryMalishev]

What

Adds ABot-World support to the engine, in two increments:

1. Model support (reviewed previously): VERSION_ABOT_WORLD detection (via the act_control_adapter.conv.weight tensor), the ActControlAdapter block (keyboard-action conditioning), Wan2.2-TI2V-5B-family loading, and a guard that rejects ABot models in the batch generate_image()/generate_video() APIs at the shared GenerationRequest::resolve() choke point (ABot is a causal interactive model, not a one-shot generator). Both capability queries report false for ABot.
2. Interactive walk (this increment): the causal walk core and a public session C API, so ABot can actually be driven:
   • src/abot_world.hpp — walk core using a recompute formulation: each denoise step re-derives attention over [reference tokens | clean history | current block] with a per-row trailing-window mask that reproduces the reference implementation's KV cache exactly (including eviction). Covers the action adapter (PixelUnshuffle'd key planes), zero-masked reference tokens, negative-time RoPE ids, per-frame timestep modulation, the 4-step distilled flow-matching schedule, and taehv pixel decoding (stateless per-block decode with a 3-latent-frame overlap standing in for the reference's streaming decoder cache).
   • include/stable-diffusion.h — sd_abot_session_{params_init,new,step,frames_free,free}: a standalone session (own DiT + taehv + scene pack, no sd_ctx_t); step(action_mask) generates the next block under the held keys (bits 0..7 = W,A,S,D,I,J,K,L) and returns its decoded RGB frames.
   • examples/abot-{parity,walk,session} — validation harnesses (static builds only; they link engine internals that the shared library does not export).

Validation (vs the PyTorch reference, F16 GGUF, fixed-scene goldens)

• Per-step parity across every attention regime — first block, all four timesteps, history recompute, window onset, deep eviction: 7/7 gates at cosine 0.9996–0.9999 (gate 0.999).
• Full-chain golden-replay walk (4 blocks × 4 steps, replaying the reference's recorded noise): latent cosine 0.9999 → 0.9931 (expected f16-vs-bf16 accumulation in a recursive rollout); decoded frames vs the reference chain: PSNR 31.3–55 dB (mean 38.9) — visually identical, and the W action drives the same forward motion.
• taehv decode vs the reference streaming decoder on identical latents: mean 46.4 dB, min 33.5 dB.
• Session C API end-to-end (Q8_0, CPU): one block → 9 RGB frames at 832×480, clean render.

Regression safety

New code is only reachable for VERSION_ABOT_WORLD models or through the new sd_abot_* entry points. Existing paths are untouched except: an optional defaulted attn_mask parameter on WanAttentionBlock::forward, and inline on two free functions in tae.hpp (ODR fix — the header is now included from more than one translation unit). Batch-generation rejection for ABot is unchanged and covered by the monorepo integration test.

[olyasir]

This fail-fast guard only covers one of the two doors into generation, and the capability layer disagrees with it.

1. generate_image() is unguarded. sd_version_is_wan() was (correctly) extended to include VERSION_ABOT_WORLD, so ABot travels through all the shared Wan latent/VAE/sampling paths used by both generate_video() and generate_image()/generate_image_internal(). But the rejection was added only here. A caller that loads an ABot GGUF and calls generate_image() runs the bidirectional/one-shot Wan recipe on a causal model with no check at all — silently corrupted output, or a raw GGML_ASSERT abort if it hits a shape assumption that was never true for ABot. That's exactly the hazard this guard exists to prevent.

2. The capability query now lies. sd_version_supports_video_generation() (line 3104) is ... || sd_version_is_wan(version) || ..., so it now reports ABot as video-capable. The example server's /vid_gen route (examples/server/routes_sdcpp.cpp) uses that query to reject unsupported models cleanly before reaching the engine — it will instead wave ABot through to this return false, bypassing the intended clean-rejection path.

Suggested fix: reject in one shared choke point instead of per-entrypoint, and keep the capability query honest:

• exclude ABot from sd_version_supports_video_generation() (and any image-capability query), e.g. ... && !sd_version_is_abot_world(version);
• move the actual rejection to a point both entrypoints pass through (GenerationRequest / SamplePlan::resolve()), or reject at new_sd_ctx() load time when the version is VERSION_ABOT_WORLD and no causal-session API was requested.

Loading still works; generation is then refused everywhere with the same message, and the server pre-screen agrees with the engine.

[DmitryMalishev]

Both points confirmed and fixed in 23d8fd5 — thanks, this was a real gap.

• Shared choke point: the rejection moved into GenerationRequest::resolve() (same pattern as validate_ideogram4_uncond_model), so generate_image() and generate_video() are both covered by one check; the per-entrypoint guard here was removed. Message now names both entrypoints.
• Honest capability queries: sd_version_supports_video_generation() and sd_version_supports_image_generation() both return false for ABot (image could not simply stay !video, or it would have flipped to true).
• script/validate_abot_world.sh grew an img_gen lane asserting the image path rejects too; verified locally against the ABot Q8_0 GGUF (detection + full load unchanged, both batch paths reject, stock-Wan lane unaffected).

On load-time rejection: kept loading permitted intentionally — downstream (tetherto/qvac#3352) exercises load on GPU CI ahead of the causal session API, and the capability queries now give front-ends the clean pre-screen.