[soltros]

A mipmapped 3072-square ASTC texture is about 48 MiB decoded, below the existing 64 MiB trimming threshold. Several such images can crowd a small GPU's working set. On Minimum with a known driver budget at most 4.5 GiB, use a 1024 dimension limit and 8 MiB threshold for sampled ASTC mip trimming.

Adopts and adapts NotRexed/Hookshot#1009 by pablitorockz, original head 9390d68afff33046a0529c3b37c23d39926d1bd0. The reconstructed import remains an ancestor. Existing Longshot texture-size validation, create/as_target signature, profiler scope, test discovery and independent README are retained. The imported unrelated answer-domain test changes were omitted.

Reuse the existing mip/view remapping and retain at least one mip. A 3072-square texture uses its 768-square level under the new default. Other presets, unknown budgets and budgets over 4.5 GiB keep existing defaults. Uncompressed/depth handling and the separate allocation-size policy are unchanged. This intentionally trades sampled-texture detail for a smaller working set; it does not promise lower total driver allocation or fix every texture-corruption report (#3/#61).

RX_VK_LOW_VRAM_MIPS=0 opts out. RX_VK_TEX_MAX_DIM still overrides dimensions. Valid RX_VK_TEX_TRIM_MB=1..1536 overrides the decoded-size threshold. Integration preserves the existing RX_VK_TEX_SKIP_MB override, including zero and values above 1536: valid TRIM_MB takes precedence; absent/invalid TRIM_MB falls back to the legacy override, then the tier default. Restart to recreate existing textures after changing the preset or environment; the current preset affects newly created textures rather than restoring existing mip levels automatically.

The authored fixture executes the production selection block with real NVN/Vulkan types and a stubbed GPU-budget query. Expanded from 12 to 22 configurations: budget boundary/just above, preset changes, opt-out, overrides and invalid values, legacy compatibility/precedence, UI-sized images, final mip and representative render/depth images. All pass with GCC and UBSan using the workflow's pinned Vulkan-Headers 74d8a6cb930c68ef617b202c3ff3c59d919e086b. The unchanged main source fails the new default selection case. No driver, game assets or gameplay were used. Upstream's historical RTX 2050 measurements remain source-author claims at the original PR, not new Longshot performance evidence. No current GPU screenshots or performance measurements were collected here.

Published full tree matches local HEAD e724e87415a8ed56591d501ac7010d164eb30fd7. Full local suite: 281 passed, 12 unchanged host/tool-dependent failures, 56 skipped; failure paths match the prior run. The new header-dependent fixture was skipped by automatic discovery in that run, and was separately executed successfully with the explicit pinned header path and UBSan as described above. Fresh hosted checks are pending. Keep draft for platform validation and review of the visual tradeoff; the broad hardware reports remain open.

Platform follow-up: the initial Linux suite passed 338/0/11; Windows passed 327 with one fixture compile failure (getenv deprecation under warnings-as-errors). The fixture now uses _CRT_SECURE_NO_WARNINGS on Windows, matching the other native fixtures and retaining -Wall/-Wextra/-Werror. All 22 cases still pass locally; fresh Windows checks are pending at 31179ad. No texture-policy code changed in this follow-up. Linux packaging has the separate Launchpad bootstrap blockage tracked by #126.

Final integrated validation

The corrected Windows fixture passes. Validation-only #127 tested this component head with #124/#125 and #126's repaired Linux bootstrap at cd33b08, tree d2336ac28805e8097b20547b13f7360cfdd2da70. Windows: 329 passed/0 failed/22 skipped; Linux: 340 passed/0 failed/11 skipped. Both packaging builds, Linux backend/player checks, Vulkan/DLSS, actionlint and focused shader jobs all passed. The conditional Minimum-preset policy and override compatibility are accepted based on these regression/build checks; this is not a new GPU performance result. #3 and #61 remain open. Final main will be compared with the integrated tested tree.