Add reasoning-pattern benchmark scaffold

benches/README.md

@@ -73,6 +73,56 @@ python benches/llamacpp_bench.py tput \
 
 Use `--json-out <path>` to save machine-readable results.
 
+## Reasoning-pattern benchmark
+
+`reasoning_bench.py` compares Direct, Best-of-N, Tree-of-Thought, and
+Graph-of-Thought through method-specific benchmark inferlets by default.
+Reference answers remain in the Python harness and are never sent to the
+model. The older `reasoning-benchmark` inferlet remains available as a
+prototype/reference with `--inferlet-mode prototype`.
+
+Build the base and method-specific inferlets:
+
+```bash
+cargo build --manifest-path inferlets/reasoning-base/Cargo.toml --target wasm32-wasip2 --release
+cargo build --manifest-path inferlets/reasoning-direct/Cargo.toml --target wasm32-wasip2 --release
+cargo build --manifest-path inferlets/reasoning-best-of-n/Cargo.toml --target wasm32-wasip2 --release
+cargo build --manifest-path inferlets/reasoning-tree-of-thought/Cargo.toml --target wasm32-wasip2 --release
+cargo build --manifest-path inferlets/reasoning-graph-of-thought/Cargo.toml --target wasm32-wasip2 --release
+```
+
+Run the bundled smoke problems:
+
+```bash
+uv --project sdk/python-server run python benches/reasoning_bench.py \
+  --driver cuda_native \
+  --model Qwen/Qwen3-0.6B \
+  --pattern all \
+  --json-out .tmp/reasoning-smoke.json
+```
+
+For GSM8K, pass a JSONL file containing the official `question` and `answer`
+fields. The harness extracts the numeric reference after GSM8K's `####`
+delimiter.
+
+`inferlets/reasoning-base` is the minimal prompt-to-completion inferlet. It
+does not extract answers, score candidates, or implement a scaling method. The
+method-specific benchmark inferlets share one internal implementation crate,
+`inferlets/reasoning-core`, so they preserve a common comparison schema:
+`pattern`, `final_response`, `final_answer`, `selected_candidate_id`,
+`candidates`, and `stats`. Use `--inferlet-dir PATTERN=PATH` to point one
+method at an external/user-submitted inferlet with the same comparison schema.
+
+RatioThink's ToT implementation is a useful reference for the next refinement
+pass. The pieces worth adopting first are explicit beam/frontier bookkeeping,
+separate `ok`/`incomplete`/`error` node states, observable scorer failures, and
+final-answer selection from the deepest surviving beam rather than blindly
+trusting the last generated level. Its source also records an important Pie
+runtime lesson: in the current host path, sibling decode from one inferlet does
+not reliably coalesce into larger GPU batches, so a memory-frugal sequential
+or coupled expansion strategy can be the right baseline until forward-pass
+deferral improves.
+
 ## Fairness defaults
 
 - vLLM prefix caching is disabled.