Bump CrispEmbed to v0.17.6

[niksedk]

Moves the pinned CrispEmbed release from v0.17.5 to v0.17.6: the seven download URLs in CrispEmbedDownloadService.cs, plus a new index-0 entry in each of the seven DownloadHashManager.CrispEmbed.* lists so the .installed.sha256 sidecar check does not offer an "update" to the version the user just installed.

What is in v0.17.6, and why none of it changes SE behaviour

It is a reranker-quality and reliability release:

• POST /rerank could abort the whole server process on quantized scoring heads — SE never calls /rerank.
• mxbai / ms-marco reranker GGUFs re-shipped with the missing pooler stage, and imatrix re-collection for the 4-bit rerankers — SE ships no rerankers.
• An env-gate sweep where FOO=0 turned a feature on (UOCR_PD=0 even segfaulted) — the unlimited_ocr engine, which none of SE's five backends use.
• CRISPEMBED_*_BENCH gates — diagnostic prints only.

The one worth checking was DS2_CROP_MODE=0, which CrispEmbedOcr sets to keep DeepSeek-OCR-2 in single-view mode. Diffing src/deepseek_ocr2.cpp between the tags: it already had its own inline strcmp(e, "0") != 0 parse in v0.17.5 and is explicitly exempted from the sweep in v0.17.6 ("different absent semantics"). So SE's request was honoured before this bump too — the measured 3–6× speedup was real, and nothing changes here.

So this is housekeeping, not a fix: no user-visible OCR change is expected.

Verification

• Hashes. Downloaded and hashed the macOS, both Linux and both Windows non-CUDA archives — all five match what is committed. The two CUDA archives (675 MB / 752 MB) use GitHub's release-asset digest field instead; that same field reproduces all seven v0.17.5 hashes already in this file, and the five it was cross-checked against here.
• Linux loader. Parsed .dynamic / .gnu.version_r out of every binary in crispembed-linux-x86_64.tar.gz: no unresolved DT_NEEDED, max GLIBC_2.27. The OpenBLAS fix behind Linux: CrispEmbed server exits with code 127 #13205 and the manylinux floor both hold.
• Output parity. GLM-OCR q4_k through crispembed-server + POST /ocr/model on EN/DE/RU/JA subtitle bitmaps returns byte-identical text and per-token confidences on v0.17.5 and v0.17.6 (only ms differs).
• Build clean; the 22 CrispEmbed tests pass.

Not re-run: the full 9-image DeepSeek-OCR-2 corpus (the model is a 2.3 GB download and the release touches nothing on that path).

🤖 Generated with Claude Code