feat: implement generic LLM-based template logic executor

[ad23b1012]

Description

This PR introduces a generic Large Language Model (LLM) executor core into template-engine. It allows templates to execute logic strictly based on their natural language text and Concerto data models without requiring an explicit TypeScript logic.ts file.

By passing runtime: "llm", the engine falls back to the new LLMLogicExecutor which bridges reasoning APIs with strictly validated Concerto schemas.

Key Features

• Generic LLM Executor Module (src/llm): Exposes the LLMLogicExecutor, responsible for evaluating request structures using natural language and model context instead of a JS compiler.
• Provider API Abstraction: Clean, native Node-18 fetch() classes for interacting with OpenAI endpoints (including Ollama, vLLM) and Anthropic Claude APIs without causing SDK dependency bloat.
• Hierarchical Config Fallbacks: Supports injecting LLMConfig configurations straight from the archive's package.json (accordproject.llm), Node environment variables, or run-time parameter overwrites.
• Strict Typed Output Validation (ResponseParser): Guarantees identical type-safety to legacy TypeScript pipelines by explicitly stripping Markdown formatting out of API responses and piping all outputs retroactively through the engine's legacy Serializer.fromJSON validation schema endpoints.
• Self-Healing Loop: If the evaluated LLM fails to output valid JSON compliant with the Concerto model, the validation error is automatically passed back to the LLM agent to retry and correct itself.

Architectural Impact

Modifications directly to the TemplateArchiveProcessor.ts init and trigger pipelines were kept intentionally minimal to prevent breaking existing infrastructure. Legacy logic execution continues normally if the evaluated compilation language remains typescript.

Testing

• ✅ Added LLMLogicExecutor.test.ts (24 new test suites covering retry injection sweeps, prompt generations, model context attachments).
• ✅ Verified all native TemplateArchiveProcessor TS evaluations continued functioning completely unaffected structurally.
• ✅ Created an offline test environment archives/latedeliveryandpenalty-llm ensuring full serialization.

---

Note for reviewers: To properly parse an llm folder archive natively via .fromDirectory(), a small upstream whitelist adjustment PR strictly permitting llm runtimes on Metadata.ts via cicero-core is expected as a follow-up.