[const-apex]

Summary

This adds two opt-in configuration options for running SwarmWorld against open-weight servers (vLLM, llama.cpp) and one diagnostic improvement.

All defaults reproduce current behaviour; the paper profiles are unaffected, and the prompt and action-schema hashes recorded in trace headers are unchanged unless a flag is set.

Changes:

llm.transport: chat_completions

(default: responses).

llama-server's /v1/responses converts requests to chat completions internally and drops text.format, so JSON comes back unconstrained (18 of 33 plans failed local validation in our test).

With this option:

• the provider posts to {base_url}/chat/completions with response_format: {type: json_schema, json_schema: {name, schema, strict}}, the same schema object the Responses path sends.
• max_tokens, temperature and reasoning_effort map one-to-one;
• usage keys are mapped back to input_tokens / output_tokens / total_tokens so traces and the compute audit stay uniform;
• finish_reason: length raises the same incomplete error;
• a leading <think>...</think> block is stripped when a server returns reasoning inline.
• llm.extra_body merges arbitrary request fields into either transport (used for chat_template_kwargs.enable_thinking: false with Qwen on vLLM).

llm.strict_id_patterns

(default false).

Small models copy the prompt's "use NONE when irrelevant" into reply_to and target_artifact_id (29 and 22 of 42 actions with Gemma 4 E4B), and the engine rejects every such action.

With the flag on:

• the three free-string ID fields get a pattern in the schema (^(message_[0-9]+)?$, ^(agent_[0-9]+)?$, ^(artifact_[0-9]+)?$), so constrained decoding cannot emit placeholders;
• the reproducibility hash includes the flag, so traces are self-labelled.
• Requires a backend that honours pattern (vLLM: guidance; xgrammar rejects it).

Diagnostics

On finish_reason: length the error string now records the length, head and tail of the truncated reply, so runaways can be classified from the trace.

Tests

Tested:

• ruff check src tests clean;
• 154 tests pass in the engine-only environment (the 25 deselected ones need the analysis extras and are untouched), including the 7 new ones;
• with the flags off, the action schema is byte-identical to the unpatched module (same SHA-256), and prompt_sha256 in trace headers equals the released traces' value.
• In use: ~50,000 model calls across Gemma 4 E4B, Qwen3-235B (llama.cpp RPC) and Qwen3.8-27B (vLLM, three quantisations), including a complete N=50 row of the paper's protocol.

Related: #1 (late state-digest divergence when replaying the released traces on revision 9); this PR does not touch replay.