fix(lcm): transcript batch with a summary message writes no summary node

[ScriptedAlchemy]

session_suite transcript_store::concurrent_full_batches_converge_without_split_brain_or_partial_writes (crates/tracedecay/tests/session_suite/transcript_store.rs:639) fails with summaries: 0 vs 1 on PR #707 after 36511ba7b unblocked the profile-scoped write: a three-message batch that includes a kind="summary" / codex_context_compacted message commits three session_messages rows but writes no lcm_summary_nodes row.

Regression in the summary-node projection off a transcript batch (global-db transcript persist → session-temporal-store summary projection). The call site was not located; the previously masking failure (git evidence in Profile scope) is fixed.

Found during the 2026-09-04 CI sweep; reported by the daemon_suite lane. Related: #843, #848.

[ScriptedAlchemy]

Not a projection regression. Transcript persistence never wrote lcm_summary_nodes; those rows come only from LCM compression (dag::insert_summary_node → publish_immutable_summary → project_canonical_summary). The failing assertion held a stale expectation, corrected in 61c8d5c. 7777cf1e2 (on PR #707's branch) pins the other half of the contract: a codex_context_compacted summary message keeps the body, kind and metadata discriminators native_summary_evidence keys on through persist_transcript_batch. Evidence: session_suite transcript_store 9 passed.