[leehack]

Summary

Reject budgeted wrapper loads before upstream entry when effective auto-fit is disabled, including explicit parameter placement, or when the default integrated-only GPU route cannot plan the budget. Return NULL with actionable sd_dart_last_error() even with logging muted. Unbudgeted loads and supported auto-fit routes remain available.

Refs #34 and 0.18.0 plan. This is a bounded wrapper restriction; raw upstream new_sd_ctx remains unsafe and existing Dart guards stay. Does not close #34 or publish an artifact.

Validation

• Normal, Metal-defined and ASAN guard controls pass; previous candidate fails the explicit-placement regression.
• Native teardown suite passes with one Linux-only skip; pinned-Dart progress stress passes all five tests; device/log tests pass.
• Prior guard head 795aa2c passed independent blocking review and actual macOS release/ASAN runtime controls. The real-built library rejects both guarded combinations before allocation; this is not full affected-device inference qualification.
• Current head a4ce93b660e47f7d4129ab8380a88bab78d8cac3, tree d13d4208c070e0dfc6c4a9496de89a8b88aea235, merges qualified main 2b4773a19354528010d131c57a6341ab64eeba97 (PR Pin Linux x64 qualification and release builders to Ubuntu 24.04 #36). Guard sources are unchanged; renewed exact-head audit accepted; all17 current-head CI checks pass. Three workflow contract tests pass.
• Positive controls cover no budget, explicit backend with empty parameter placement, discrete default, and recovery after rejection.

Remaining gates

All17 current-head CI checks pass; renewed independent audit accepted the byte-identical guard diff against the new base. Earlier Linux Vulkan CI failed after floating runner/toolchain drift; PR #36 pins builders and qualification to Ubuntu 24.04 without relaxing thresholds. The old run is not current-head evidence. Affected-device qualification remains required before publication. The explicit-placement variant was established through upstream planner code and a regression control, not a new real-model crash reproduction. No wider GPU support claim.

[leehack]

Additional exact-head local artifact checks at 795aa2c: macOS arm64 Release build succeeded and validate_artifacts.py macos-arm64 passed. Built library SHA-256 3e8568483468a6e0bf7a30a01b12b88d65bcc1c7fd614631fad3cdf754d120aa.

The maintained built-runtime teardown matrix passed (8 tests, 4 skips) with SD_REQUIRE_RUNTIME=1, pinned Dart and SD_REQUIRE_METAL_GENERATION=1; its untracked control reproduced the expected ggml-Metal exit abort. A linked C++ probe using the SD-Turbo Q8 model path safely rejected both fit-off and explicit-placement budget configurations with actionable errors. The guard rejects before model allocation, so this is real-library boundary evidence, not image-generation qualification. Model SHA-256 d50be7655f0a554cf8041c145d88b210bd5f3c545423119dee62ae08cae51580.

AddressSanitizer runtime matrix is still running. Linux runtime and broader affected-device/model qualification remain open; no publication readiness claim.

[leehack]

The AddressSanitizer built-runtime teardown matrix also passed on the same source head: 8 tests, 5 skips, required Metal generation enabled. Its untracked control again produced the expected Metal exit abort. Linux runtime and broader affected-device/model gates remain outstanding.

[leehack]

At head a4ce93b660e47f7d4129ab8380a88bab78d8cac3, fifteen jobs passed. The iOS x64 simulator job was cancelled before any step ran: GitHub reports that no hosted runner acquired it after multiple attempts, with macOS arm64 capacity warnings. One targeted retry of that job was dispatched; its dependent XCFramework check still needs to run. Merge remains gated on those results.