Pin llama.cpp d1747cb: Q2_K weights on HRX0 (llama.cpp #49)

[bong-water-water-bong]

Moves third_party/llama.cpp from b8d587e to d1747cb. The only change is llama.cpp #49, which gives Q2_K weights a kernel on HRX0.

Why it matters. Q2_K had no matmul on HRX0. At load time llama.cpp tests each weight with a 512-token matmul to decide where it lives; that test failed, so every Q2_K weight went to the CPU_REPACK buffer, and each use split the graph. Every Unsloth UD GGUF carries some Q2_K.

What #49 does

• Adds Q2_K to the shared dequantizer, which covers prompt WMMA, generic decode and GET_ROWS.
• Adds Q2_K lane functions to the K-quant decode kernels, for 1 token and for 2-8 tokens.

Measured on strixhalo (HRX0), Qwen3-4B requantized Q8_0 → Q2_K

• Placement: 709 MiB in CPU_REPACK with 217 graph splits before; all weights on HRX0 with 73 splits after.
• Speed: pp512 686 → 1287-1325 tok/s. tg128 47.0 → 74.4-74.8 tok/s; the spread of ±14 is likely first-run kernel compiles.
• Accuracy: KLD against the Q8_0 model is 1.1045 on the prefill path and 1.1031 on decode (-b 1), against 1.1140 on CPU and 1.1224 for the previous HRX with Q2_K on the CPU.
• Tests: test-backend-ops -b HRX0 passes 929/929; MUL_MAT q2_K 11/11.

Also in this PR:

• registry/architectures.json regenerated for the pin. tools/registry_build.py --check-pins passes, and so does tools/check_pins.py origin/main (ahead).
• docs/hrx.md: the sub-4-bit note now says what still falls back to the CPU (IQ2_XS/XXS, IQ1_S/M) and what context7.json: claim the Context7 library #49 changed.

🤖 Generated with Claude Code

[context7]

Docs7 for 1bit-monster/engine

Result	Status	Action
Deployment	➖ Not used	—
Content review	➖ Did not run. This site has no agent runs available this month. Wait for the monthly reset or check your Docs7 plan.	—

Commit f71a758

[bong-water-water-bong]

Review (standing in for PR-Agent while coder-llm is paused)

• Pin: b8d587e..d1747cb is exactly fork context7.json: claim the Context7 library #49 (738460e, the review nit ef779b8, and the merge). I reviewed context7.json: claim the Context7 library #49 line by line against dequantize_row_q2_K; see my comment there.
• Registry: llama.cpp (hrx) records d1747cb, so registry_pins passes. The counts are unchanged, as expected for a weight-format change.
• docs/hrx.md: the sub-4-bit note is accurate. IQ2_XS, IQ2_XXS and IQ1_S/IQ1_M are still on CPU, which is why UD files split. Q2_K moved to HRX0 with the measured numbers.
• No HRX re-check of other models needed: the change only adds a format branch (12) that existing formats never take. context7.json: claim the Context7 library #49's test-backend-ops 929/929 and its KLD numbers cover Q2_K itself.

Good to merge once build is green. pr_agent failing is infra while the model is paused.