feat: wire real LLM runner via third-one or OpenAI-compatible API

openspec/changes/bootstrap-job-tutor-platform/specs/tutor-workflows/spec.md

@@ -56,3 +56,30 @@ boundary.
 - **WHEN** it invokes the workflow layer
 - **THEN** it calls a typed Go interface
 - **AND** does not mutate product state by parsing freeform shell output.
+
+### Requirement: LLM runner calls OpenAI-compatible API
+
+The system SHALL provide an LLM-based workflow runner that implements the
+Runner interface by calling an OpenAI-compatible chat completions API when
+TUTOR_LLM_API_KEY is configured.
+
+#### Scenario: grader uses LLM when configured
+
+- **GIVEN** TUTOR_LLM_API_KEY and TUTOR_LLM_ENDPOINT are set
+- **WHEN** the server starts
+- **THEN** an LLMRunner wraps the configured model
+- **AND** GradeInterviewAnswer calls the LLM with a structured grading prompt
+- **AND** the response is parsed into the typed GradedAnswer contract.
+
+#### Scenario: memory extraction uses LLM when configured
+
+- **GIVEN** an LLM runner is active
+- **WHEN** ExtractLearningMemory is called with a graded answer
+- **THEN** the LLM produces MemoryUpdateCandidate with concept mastery, misconception, intervention, and review schedule updates.
+
+#### Scenario: falls back to stub when unconfigured
+
+- **GIVEN** TUTOR_LLM_API_KEY is empty
+- **WHEN** the server starts
+- **THEN** a StubRunner is used
+- **AND** grading and memory extraction produce deterministic stub output.