Shared builder for the canonical turn event — the substrate-level record of one driver turn, emitted identically by both turn drivers (plan D1):

• PtcRunner.Session turns (external LLM drives via MCP / mix ptc.repl), and
• the PtcRunner.SubAgent loop (the internal loop drives the LLM).

Both drivers build their turn record through build/1 so the top-level shape is identical and queryable through the same PtcRunner.TraceLog.Analyzer calls, regardless of which driver produced it. Driver-specific richness lives in the data bag; the top-level fields never diverge.

The map is a v2-flat-compatible envelope (event: "turn", schema_version: 2) so it rides the existing PtcRunner.TraceLog JSONL/in-memory sinks. The sink stamps trace_id, timestamp, and seq; this builder never sets them.

Top-level fields

schema_version, event ("turn"),
driver ("session" | "sub_agent"),
session_id, agent_id, agent_name,
turn, attempt, committed, status,
duration_ms, input_tokens, output_tokens, total_tokens,
data

turn is the committed-state counter (advances only when an attempt commits); attempt is the monotonic per-attempt counter (advances on every attempt, including failed ones); committed flags whether this attempt advanced committed state. Failed/parse-error/budget-stop attempts are recorded with committed: false so wasted work is visible without mutating driver state (plan P2 notes).

data bag

program, raw_response, result_preview, prints, memory_diff,
tool_calls, limits_hit, preludes, fail, turn_type

Per-driver fields that don't apply are nil/empty. raw_response carries what the driver's LLM generated when there is no parsed program (SubAgent parse/text-mode failures); it is nil for session turns and normal turns. The whole bag is run through PtcRunner.TraceLog.Event.sanitize/1, which bounds large strings/lists/maps — that is where memory-diff values and prints get their byte bounds.