#2610 setup: pin Swift 1.5 at 99bb8f7f (2026-10-07), the first revision with its IQ3_S files

本文

**What it changes.** One line in `setup.py`'s `HF_REVISIONS`: `ukisai/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-GGUF` moves from `b22d729e` (2026-09-24) to `99bb8f7f95c7aa7b24a36a7786a4f657b30f5d3d` (2026-10-07). The old revision has no IQ3_S shards, so every Swift IQ3_S install (setup's recommended Swift size since 0.1.44) prints "not at the pinned revision any more" and downloads the repository's current files unpinned (seen in this box's own setup log). Open PR #1610 (by @xtremexp) carries the same bump among its other changes; this PR is only that line, credited to it.

**Verification** (Hugging Face API, `tree/<rev>?recursive=true&expand=true`, LFS sha256 and size at both revisions):

| File | Size (bytes) | LFS sha256 | b22d729e -> 99bb8f7f |
| --- | ---: | --- | --- |
| Swift-…-IQ2_XS-00001-of-00002.gguf | 39,788,473,344 | ca3b302d3d1eb6ed… | same |
| Swift-…-IQ2_XS-00002-of-00002.gguf | 28,363,693,824 | e4996a6de18fd49c… | same |
| Swift-…-IQ3_XXS-00001-of-00002.gguf | 39,785,790,560 | 3bddaa667c750f63… | same |
| Swift-…-IQ3_XXS-00002-of-00002.gguf | 36,180,282,560 | b0b15f782af71eb4… | same |
| Swift-…-Q2_0-00001-of-00002.gguf | 39,799,117,984 | 79e2a3873f1b05df… | same |
| Swift-…-Q2_0-00002-of-00002.gguf | 26,750,834,816 | ef3bb04fb4e14f04… | same |
| imatrix-swiftfn-v1mix.gguf | 580,038,816 | 769098d84e38baa2… | same |
| mmproj-Swift-Qwen3.8-Flash-Next-BF16.gguf | 907,543,520 | cd1140f4abba943c… | same |
| Swift-…-IQ3_S-00001-of-00002.gguf | 44,900,816,736 | 46cb3996cd1a17fe… | new |
| Swift-…-IQ3_S-00002-of-00002.gguf | 38,842,920,384 | 70ce6fb48a94fe34… | new |

Nothing was removed; only metadata setup does not download differs (README, SHA256SUMS, manifests, evaluation files, the new IQ3_S recovery and allocation files).

**What it leaves alone.** Every other pin and all code. Installed Swift files are the same bytes at both revisions, and an install's finish marks keep it from being fetched again.

**How it was checked.** The table above (re-checked independently by a sub-agent reviewer, who also read how `hf()` and the "not at the pinned revision" fallback use the pin; no test asserts the old revision). `tools/test_setup_*.py`: 38 of 40 files pass; `test_setup_engine_hash.py` (2 failures) and `test_setup_golden.py` (8 failures) fail the same way on `main` on this Linux box (Ubuntu, Python 3.14).

🤖 Generated with [Claude Code](https://claude.com/claude-code)