[stikves]

This PR makes the batched engine from #335 usable from llm-runner.

What this adds

• Engine selection: a dynamic graph whose main function takes an attn_mask input now auto-resolves to coreai-sequential, the only engine that fills the mask. Before, auto-detection picked the pipelined engine, which fails on these bundles with an input-count error. An explicit coreai-pipelined override on such a graph is rejected, and incompatible-override errors now include the reason.
• llm-runner --batch-prototype N: a diagnostic that runs the prompt as N identical rows through lockstepBatchedGenerate and prints each row. Row 0 is greedy. The other rows sample at --temperature (0.8 if that is 0) with a fixed seed per row. The row count is checked before the model loads, and a count above the bundle's max_batch_size is rejected.

Scope and behavior

• Bundles without an attn_mask input select the same engine as before.
• --batch-prototype is off by default and adds no public API.

Validation

• Build and swift format lint --strict are clean. The full Swift test suite passes, apart from the existing Top-K latency timing test, which can fail when the whole suite runs in parallel and passes on its own. New tests cover the attn_mask variant resolution and the rejected pipelined override.
• On a dynamic-batch Qwen3-0.6B bundle exported with --attention-mask batched --dynamic-batch-size 4:
  • the default variant now runs and gives the same greedy output as an explicit coreai-sequential,
  • --batch-prototype 1, 2 and 4 all run; row 0 is identical at every N and matches a normal greedy run, and a rerun reproduces every row,
  • a row count of 5, a count of 0, and a pipelined override are each rejected with a clear message.
• Greedy output with the option off is byte-identical to main on the default, coreai-sequential and coreai-pipelined engines, for a fresh dynamic export and an older dynamic bundle.
• A static-shape bundle still selects the static-shape engine and gives the same greedy output as before.