Adds first-class support for Retell's function (tool-call) nodes and branch (logic-router alias) nodes, then centralizes transition evaluation across every node type. Tool execution itself stays out of scope and is tracked at #51.

Summary

• New NodeType.FUNCTION with importer (function → FUNCTION), exporter (FUNCTION → function), engine handler (warning + delegate, no HTTP call), and is_function_node() helper.
• branch aliased to NodeType.LOGIC in the Retell importer. Previously fell through to NodeType.CONVERSATION silently — a real correctness bug on newer Retell CF exports.
• Centralized _evaluate_transitions dispatcher in voicetest/engine/conversation.py. Every node type (logic / extract / function / conversation) routes through one method that dispatches per condition.type:
  • equation → deterministic via evaluate_equation
  • llm_prompt → fires when caller-supplied llm_decision.transition_to matches the target
  • tool_call → skipped with a per-transition warning (until Tool execution for function nodes (Retell CF function, etc.) #51)
  • always → fallback (conversation nodes opt out — always edges still fire post-response)
• Conversation-node equation transitions are now evaluated deterministically instead of LLM-reasoned-as-text. The objectives_complete gate still applies to all transition types.
• Broad comment cleanup. Stripped WHAT-explaining comments and multi-paragraph docstrings across 91 files (~1800 net lines removed). WHY comments preserved (DuckDB rationale, retry/idempotency contract, Retell field quirks, etc.). New "Add no comments by default" section in docs/development.md.
• DRY fix in AgentService: load_graph and get_graph_with_etag shared a "linked-file vs DB" pattern — extracted to _load_graph_payload.

Behavior changes worth flagging

• branch Retell nodes now route as LOGIC. Previously they silently became conversation nodes — anyone relying on that incorrect behavior will see a routing shift. Considered a bug fix, not a regression.
• Conversation-node equation transitions on extracted variables now fire deterministically. Previously the LLM was asked to "reason about" the equation as text; in practice this was unreliable. Equation conditions on tool_result-style variables still fall through to else_edge (variable absent → equation evaluates False).
• tool_call-typed transitions (Telnyx/LiveKit handoff shape) are now skipped with a warning instead of being silently dropped on logic/extract nodes or LLM-reasoned on conversation nodes.

Out of scope

• Real tool execution for function nodes — tracked at Tool execution for function nodes (Retell CF function, etc.) #51 (mock-mode via TestCase.tool_mocks, live HTTP execution behind a flag, result-driven branching).
• Validation that LLM-picked conversation transitions were actually offered — pre-existing risk that isn't introduced by this PR. Filed as Conversation-node transitions trust LLM-picked targets that weren't offered #53.

Test coverage

• New TestRetellImporterFunctionNode and TestRetellImporterNodeTypeAliases in tests/unit/importers/test_retell.py.
• New TestFunctionNodeExport in tests/unit/exporters/test_retell_cf.py, with import → export → re-import round-trip.
• New TestFunctionNodeHandling and TestCentralizedTransitionDispatch in tests/unit/engine/test_conversation.py. The dispatch tests cover: function-node equation routing on pre-existing variables, tool_call warning + fallthrough, conversation-node equation firing deterministically over LLM pick.
• New NodeType.FUNCTION round-trip test in tests/unit/models/test_agent.py.
• New tests/fixtures/retell/sample_config_function_node.json minimal CF fixture.

Stats

13 first-pass files + the centralization + the broad sweep:

• +664 / −2482 across 91 files (mostly comment removal)
• 1647 unit tests passing, ruff clean

Issues touched

• Tool execution for function nodes (Retell CF function, etc.) #51 — Tool execution for function nodes (referenced from engine warning + docs)
• Conversation-node transitions trust LLM-picked targets that weren't offered #53 — Conversation-node transition validation (filed alongside)