Validate Gemma 4 SWA long-context correctness against llama.cpp (cross-tool oracle)

[pekkah]

Follow-up to #164 (commit 48a5622), which fixed a latent bug: Gemma 4 SWA KV caches were window-sized but indexed absolutely, so output was silently wrong beyond the 512-token sliding window. The fix (a real pos % cacheSize ring) is validated internally — chunked-vs-per-token parity on prompts whose final-token window spans a chunk boundary (so the observed logit depends on cross-chunk + ring-wrapped reads), plus a >5K-token coherence run.

What's missing: an independent cross-tool check vs llama.cpp at long context. It was skipped during #164 because llama-cli hung in interactive mode and the chat-template token-parity matching is a known rabbit hole (see CLAUDE/memory).

Task: greedy-decode a >4096-token prompt on both sharpi (sharpi-cli.exe -g -1) and C:\p\llama.cpp-cuda\bin\llama-cli.exe with matched tokenization (verify token ids via llama-tokenize, drive llama-cli non-interactively — pipe EOF / -no-cnv confirmed effective), and compare the first ~20 generated tokens. Document the result. Low priority — internal oracle already covers the regression; this is defense-in-depth for a correctness fix.

[pekkah]

Oracle feasibility check (2026-06-22): the bundled tools/llama.cpp (build b9444-cpu) does load and run gemma4-v2-Q4_K_M.gguf (gemma4 text arch is supported — model loads, prompt processes, no "unknown architecture"). So a cross-tool oracle is viable in principle.

Blocker for autonomous validation: this llama.cpp build is CPU-only (no ggml-cuda.dll; all ggml-cpu-*). The issue is about long-context SWA correctness, so the oracle run needs a large context — which on a CPU-only build is sustained heavy CPU inference. Per the dev-environment constraint (avoid heavy CPU model runs / fan noise unless explicitly requested), this can't be driven autonomously.

To proceed, either: (a) drop a CUDA-enabled llama.cpp build into tools/llama.cpp (then the long-context oracle runs on GPU), or (b) explicitly OK a CPU llama.cpp long-context run. Leaving open.