Match Qwen reasoning history across effort changes

docs/issue117-reasoning-history/product.output.log

@@ -0,0 +1,13 @@
+
+running 1 test
+test engine::metal::qwen_mtplx::owner::issue117_product::issue117_reasoning_history_cache ... ds4: Metal device Apple M5 Max, 128.00 GiB RAM
+ds4: Metal 4 tensor API enabled for Tensor kernels
+ds4: drift-patch flags hc_stable=on norm_unify=on kv_raw_f32=off rope_exp2_log2=off math_safe=off tensor_matmul=on
+ds4: Qwen reference kernels: Powered by MTPLX — https://github.com/youssofal/mtplx
+[qwen4_exp] ngram table: streamed from SSD (file-backed pages, hot cache 1023M, prefetch on) — resident only at >=160G RAM or MTPLX_NGRAM_RESIDENT=1
+[mtplx] verify SDPA head-chunked: q_len=18 heads_per_chunk=1 chunks=24
+{"case":"reasoning_history_cache","legacy_history_tokens":40,"direct_prompt_tokens":59,"live_old_miss":true,"disk_old_miss":true,"cold_state_exact":true,"warm_cached":61,"disk_cached":61,"warm_disk_state_exact":true}
+ok
+
+test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 560 filtered out; finished in 28.37s
+