Card correction: llama.cpp gained --moe-cache-mib (2026-10-07); restate engine distinction with cited numbers

Thanks @Green-Sky for the flag. Upstream now has an LRU expert cache; the card says so and cites the third-party 16GB-tier comparison instead of the stale no-cache framing.

README.md

@@ -56,7 +56,7 @@ We also benchmarked [mradermacher's static Q2_K GGUF](https://huggingface.co/mra
 | prefill | 800 t/s | 22.6 t/s |
 | fast QA gate | 10/10, direct answers | 5/10 at short budget, 9/10 given 400 tokens — right answers buried under rambling |
 
-llama.cpp has no routing-aware expert cache: the pool is either fully resident (doesn't fit) or fully streamed over PCIe (~4 t/s). That gap is the engine, not the bits. For a llama.cpp-native path of this model, mradermacher's repo is the right link — this pack needs the Strata engine above.
+Upstream llama.cpp gained an expert cache on 2026-10-07 (#29887, `--moe-cache-mib`) — a real feature in the right direction, and it is not what this paragraph was written against. The distinction that still matters is reactive vs routing-aware: LRU pays the PCIe miss on the current token; Strata's cache predicts from the router before the miss (measured 85.9% hit). Third-party comparison on a 16 GB card with a 2-bit Flash-Next pack: llama.cpp + 8 GB expert cache 49.4 t/s vs Strata 85.8 t/s, prefill 206 vs 2,320 t/s (local-llm-lab, 2026-10-16). That gap is the engine, not the bits. For a llama.cpp-native path of this model, mradermacher's repo is the right link — this pack needs the Strata engine above (TQ2_T id 144 is not an upstream type).
 
 ## Is it still smart?
 
@@ -67,7 +67,7 @@ llama.cpp has no routing-aware expert cache: the pool is either fully resident (
 | Official Q4_K_M (4.8 bpw) | 78.8 GB | 98.2% (161/164) | not run | 68–80 t/s |
 | Community Q2_K, llama.cpp (2.56 bpw) | 56.3 GB | fast-gate 5/10 short-budget, 9/10 with room | — | 3.8–4.5 t/s |
 
-Same model, same harness (`boxy-fastcoder-v1`, official prompts + asserts, 2026-11-03). The requant at less than half the expert precision costs ~3 points on the coder bench and keeps 120 t/s — while the same-size llama.cpp 2-bit quant rambles and crawls at 4 t/s because it has no routing-aware hot cache.
+Same model, same harness (`boxy-fastcoder-v1`, official prompts + asserts, 2026-11-03). The requant at less than half the expert precision costs ~3 points on the coder bench and keeps 120 t/s — while the same-size llama.cpp 2-bit quant rambles and crawls at 4 t/s on plain offload (an 8 GB `--moe-cache-mib` lifts llama.cpp 2-bit into the 40–50 t/s class; see the engine note above).
 
 - **Groot is a reasoning model.** Answers arrive only after a thinking trace; stacks with short output budgets or code stop-words truncate it mid-thought and score zero. If your eval said "Groot is dumb", check whether it read `reasoning_content` and gave the model room to finish — that explains a large share of such reports.
 - Honest fine print: Groot's real failure mode is occasional reasoning-loop degeneration (the trace collapses into a repetition spiral and never emits code). 6 of this pack's 8 HumanEval misses are that mode, counted, not hidden.