Internal working engineering record - kept for reference, not a user-facing guide.

Parsers are adapters producing NormalizedTranscriptBatch; statement building lives behind the seam

axctl's five harness parsers (claude, codex, pi, opencode, cursor) no longer compose their own dual-write SurrealQL. Each parser is an adapter whose job ends at a NormalizedTranscriptBatch - providers, sessions, events, Turns, Tool Calls, tool-file evidence, parent edges, synthetic skill invocations, plan snapshots, compactions - built by a module-private to<Provider>NormalizedBatch function. buildNormalizedTranscriptStatements (with a BuildNormalizedTranscriptStatementsOptions bag; clearExisting defaults true, threaded per-batch for streaming parsers like codex) owns statement emission: the dual write to provider events (agent_*) and normalized records (session/turn/tool_call), idempotence clears, and edge wiring. This is within-stage normalization; ADR-0006's stage contract (typed stats, DB as the inter-stage data contract) is untouched.

Provider-specific writes that the seam deliberately does NOT absorb stay as extras appended outside the batch: per-provider token-usage rows, claude hook evidence, and claude's effectful relateInvocations (routing real-skill edges through the synthetic-skill builder would clobber real skill rows). The rule is "normalized batch + provider extras", not "everything through one type".

Conversions were proven behavior-preserving with a statement-parity harness: diffStatementSets (order-insensitive, multiplicity-respecting multiset diff) between the legacy builder and the seam path had to come back empty on rich fixtures, modulo a five-entry documented delta ledger (D1 invoked SET order canonicalized, D2 agent_event: NONE key omitted when null, D3 ordering covered by idempotent UPSERT/RELATE semantics, D4 cosmetic, D5 claude's seven write spans collapsing into one normalizedBatch span). After conversion the legacy builders were deleted and the parity tests converted to golden assertions. Byte equality modulo the ledger is the contract: a non-empty delta means the adapter is wrong - never loosen the harness.

Consequence: a sixth harness implements raw-format extraction and one to*NormalizedBatch function and gets the dual write, idempotence, and edge wiring for free; a schema change to turn/tool-call statements is edited once in the seam, not five times; and parser behavior changes are detectable as golden statement diffs rather than DB archaeology.