Pin llama.cpp d60cc4f: PrismML ternary types native on HRX, graph program cache cap

[bong-water-water-bong]

llama.cpp fork since bd5b297:

• NPU lane: fresh hw_context per generate(), fixes the idle timeout #62: PrismML's PQ2_0 / PTQ1_0 as native ggml types (142 / 143) with HRX decode on the K-quant
  kernels; Q1_0 decode moves onto them. Ternary-Bonsai-2-27B (balanced mode, llama-bench -fa 1):
  PTQ1_0 5.53 GiB 14.3 tok/s, PQ2_0 6.70 GiB 15.8 tok/s, vs 14.13 GiB / 15.3 for the Q4_0 copy;
  KLD vs CPU 0.000129 for all three. CPU decode bit-identical to PrismML's build on every tensor of
  both 27B files. test-backend-ops -b HRX0 1020/1020.
• pr-agent: a bot comment no longer cancels a review #63: GGML_HRX_GRAPH_PROGRAM_CACHE (default 64) bounds HRX server memory: ZAYA1-8B over 40
  varying-length requests peaks at 8.7 GiB instead of growing past 24.6 GiB; answers identical.

serve: a file in PrismML's ternary types goes to HRX with --device auto, rotated (prism.hadamard)
or not (Ternary-Bonsai-1.7B); another --device is refused with the converter's name. Smoke test on
strixhalo: 1bit serve -m Ternary-Bonsai-2-27B-PTQ1_0.gguf routed to HRX and answered "Paris" 3/3.
tests/prism_route.sh covers the new routes; ctest 19/19 (build without HRX).

Docs: docs/hrx.md Ternary Bonsai section (native types, table), the cache cap section, "Our patches".
Registry regenerated (no mapping changes).

🤖 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 624cb7b