fix(model cache): credit the allocator's reclaimable reserve in the VRAM budget

[lstein]

Problem

_get_vram_available budgeted VRAM on driver-free memory alone (mem_get_info), so blocks the torch caching allocator holds but is not using — freed weights and dead activations from earlier pipeline stages, whenever an empty_cache was skipped (several paths skip it deliberately) — counted as consumed. On big jobs the budget under-reported by gigabytes and a model that would have fit was partial-loaded down to its minimum weight set, while the allocator happily reused the "missing" memory for activations.

Observed in the wild on a 48 GB W7900 running a MiniMax H3 Ref2VA job: the 20 GB transformer loaded at 0.0% VRAM residency ("Loading 1.14 MB into VRAM, but only −13113.5 MB were requested") and streamed every block from RAM per step. Reproduced deterministically at small scale on CUDA: after freeing 1 GiB without an empty_cache, _get_vram_available reported 128 MB when 1,152 MB was obtainable.

Fix

• Credit reserved − allocated − inactive_split_bytes in the CUDA/ROCm and dedicated-XPU branches: the reserve is reusable directly, and empty_cache() returns whole unoccupied segments to the driver. Intra-segment fragmentation slack stays excluded. Side effect: _offload_unlocked_models now sees freed bytes immediately and stops evicting once the budget suffices, instead of always dumping every unlocked model.
• Withheld entirely under expandable-segments mode (parsed from PYTORCH_ALLOC_CONF / PYTORCH_CUDA_ALLOC_CONF / PYTORCH_HIP_ALLOC_CONF, which is where InvokeAI's own pytorch_cuda_alloc_conf setting lands before torch import): there, holes are not inactive splits, empty_cache reclaims nothing, and adversarial review reproduced a hard OOM on a load the credited budget claimed would fit.
• New first-pass WARNING when the budget stays negative after offloading, naming every model still resident in VRAM (full cache key, size, lock state) and the effective working-memory reservation — until now nothing in the default log said what was occupying the device when a load fell back to minimum weights.

Verification

• New CUDA tests (run on an RTX 5060 Ti): the invisible-reserve reproduction (128 → 1152 MB), the offload-under-working-memory-pressure guard, the diagnostic warning naming locked residents, and the expandable-segments withhold. Full model-cache suite: 336 passed.
• Adversarial fresh-context review ran before push; its confirmed findings (expandable-segments over-credit → hard OOM; warning printing 0 MB for a defaulted reservation; ambiguous submodel keys in the resident list) are fixed in the second commit. It also verified: no double counting across all _get_vram_available call sites, the cudaMallocAsync backend credits honestly, the native allocator's release-and-retry makes whole-segment credit safe, no non-default-stream usage in any load path, and the paced-lock warning fires exactly once per stream.

🤖 Generated with Claude Code

https://claude.ai/code/session_01S4B5exWsbC2z2Uu2tA167A