[leehack]

Moves stable-diffusion.cpp to master-982-0518b34 (ggml dc998771) and keeps Metal builds correct under upstream's new batched_cfg=true default. Refs #20. Not a release; full record in docs/upstream_982_qualification.md.

Consumer ABI delta

	v0.2.0-4	This branch
sizeof(sd_ctx_params_t)	320	336
batched_cfg (bool)	absent	offset 289, former padding; init true
flash_attn_mode, conv_direct_mode (const char*)	absent	offsets 320, 328; init null (legacy bool route)
Other sd_ctx_params_t members	unchanged offsets	unchanged
sd_img_gen_params_t / sd_vid_gen_params_t / sd_sample_params_t	552 / 584 / 96	unchanged
convert_with_components	11 args	13 args: appends const sd_lora_t* loras, int lora_count
New exports		sd_get_preview_info, sd_ctx_supports_ref_image_conditioning

sd_ctx_params_init writes past an old 320-byte allocation: llamadart must regenerate bindings and map batched_cfg and both mode strings explicitly before loading this runtime.

Attention/convolution modes. A null mode honors a legacy flag only when it is true. With the flags false (the init default and what llamadart writes today) upstream runs its auto probes: on CPU, auto enables flash attention, so the image changes from v0.2.0-4 (4ca7bc91 → bf06aed2). On Metal, auto is not seed-reproducible: flash attention is picked by a timing probe at load (the auditor saw it on in 2 of 10 loads). To keep v0.2.0-4 behavior llamadart must always pass both modes:

v0.2.0-4 flags	flash_attn_mode
flash_attn	"off,te=on,clip_vision=on,vae=on,diffusion=on"
only diffusion_flash_attn	"off,diffusion=on"
neither	"off"

conv_direct_mode = "off" + ",diffusion=on,control_net=on" for diffusion_conv_direct + ",vae=on" for vae_conv_direct. Verified: candidate CPU "off" = v0.2.0-4 flags-false image and "off,diffusion=on" = v0.2.0-4 diffusion_flash_attn image. On MTL0, three loads of each explicit mode gave identical images and logged no auto probe.

Metal batched CFG (leejet#2149)

ggml-metal GROUP_NORM still writes only batch item 0, so batched CFG returns a wrong image as success. sd_dart_new_sd_ctx now loads a copy with batched_cfg=false when built with Metal (SD_DART_METAL from SD_METAL); other builds pass the flag through. Raw new_sd_ctx is unchanged.

Control (Apple M4 Max)	Result
Model-free GROUP_NORM [8,8,64,2], sentinel 123456	CPU all written; MTL0 batch-2 item 1 4096/4096 unchanged
SD-Turbo F16 + TAESD 128², seed 7, CFG 2, raw MTL0	unbatched mean 129.74; batched range 71..147, 46,710/49,152 bytes off by >16
Same, CPU	batched == unbatched == published v0.2.0-4 CPU image, byte-identical
Guarded sd_dart_new_sd_ctx, batched_cfg=true, MTL0	byte-identical to unbatched Metal
Published v0.2.0-4 vs candidate unbatched, MTL0	mean abs diff 0.15, max 5
batched-cfg stand-in, -DSD_DART_METAL=1	passes; fails with the guard removed (mutation check)

Carry inventory

No carry retired. Two carries the pin lacks: Metal unbatched CFG (wrapper, this PR) and patches/pickle_line_span.patch for #23 (merged in #26 and brought in by merging main). The patch applies to 0518b34 (pickle_io.cpp is byte-identical to the baseline) and test_pickle_span passes. Kept, with reasons in the doc: progress ring, log/error ring and spin abandonment, context registry and guarded calls, Apple hidden __cxa_atexit, bounded teardown waits, Linux owned-static retention and GCC 13 floor, static-registry device queries, GPU status mapping, free-memory clamp (ggml 65c60b2 overlaps but is unqualified on real Vulkan), Metal fail-loud (#21), Windows short build path + EP_PREFIX, export allowlist, MSVC generator, Android 16 KiB alignment and CPU-only, CPU desktop payload, build policy, Apple minimum OS and privacy manifest. The submodule itself stays unpatched.

Upstream fixes adopted: 8c442fb (#2144) UTF-8 model paths on Windows, now smoke-tested on every smoke target (converts between files in a non-ASCII directory); 60423ab (#2124) / b7a74e6 (#2119) memory reclamation, unqualified (no Vulkan iGPU).

Evidence (macOS arm64, guarded build with #26's patch, dylib e4e1a2be, pinned Dart 3.47.1 SDK)

Check	Result
validate_artifacts.py macos-arm64	pass
build.py wrapper-definitions check (new)	SD_DART_METAL=1; mutation (definition removed from CMakeLists) stops the build after configure
smoke_test.py macos-arm64 --expect-device MTL	pass, incl. non-ASCII path conversion
unittest discover -s tests (SD_REQUIRE_DART/RUNTIME/METAL_GENERATION, 3 runs)	64 tests, pass, 7 skipped (Linux-only, real model, libllamadart)
test_exit_teardown_runtime.py, ASan, Metal generation	pass, 5 skipped
Metal guard control after the merge	guarded batched = unbatched (7f798ea8); raw batched still defective
Same with SD-Turbo Q4_0 real model, release + Dart VM harness	pass, 3 skipped; untracked-context control aborts in ggml-metal as before
Same, real model, ASan	pass, 4 skipped

The real-model rows ran before the merge; the merged build differs only in the patched legacy pickle reader.

Evidence logs and probes: llamadart-evidence/sd-982-pr/ on the qualifying Mac (not committed).

CI: every build job now fails if stable-diffusion was configured without the SD_DART_METAL / SD_DART_GPU_BACKEND value its target needs. The macOS and Linux smoke build jobs run test_runtime_api.py against the built runtime with SD_REQUIRE_RUNTIME=1, so a missing runtime fails instead of skipping.

Coordination: SD PR#26 landed first (aa1f6d6). This branch merges main (merge commit 0a4d0ca), and the doc and README list both carries.

Unqualified

Linux cooperative native/Dart teardown (CI covers CPU/lavapipe only), Windows (CI smoke only), real Vulkan memory/convolution/batched CFG, FP8 CPU, iOS devices and full XCFramework/signing audit, Android devices. CPU parity is not GPU evidence. SD#23 (legacy pickle >2 GiB) is out of scope.

High-risk QA

Native runtime and exit-teardown change: needs an exact-head independent review before ready. Draft until then.