Notice: one-way migration of Session transcripts to the RuntimeEvent ledger

[Astro-Han]

This is a notice, not a design thread. The technical discussion lives in #4791 and should stay there.

What is changing

Ordinary execution history is currently written twice: to session_messages as StoredMessage, and to runtime_events as RuntimeEvent. #4791 removes the double write and makes RuntimeEvent the single authority. StoredMessage remains as an API and UI projection type only.

Existing local Sessions are converted to RuntimeEvents once, during the upgrade, by an extension of the Session Importer. The read path, recovery, paging, unread markers, goal context, subagent summaries and the Agent Graph consumers all move in the same PR, so the repository never carries a half-migrated state.

Why this needs a notice

The migration is not reversible. Once it runs, downgrading to a build from before the cutover cannot read its own data. The compatibility epoch covers the protocol, not the data.

The conversion is a total function: every legacy row produces something, and anything that cannot be mapped to a structured event becomes a text event carrying its rendered content with legacy provenance. No Session and no line of conversation is dropped. Three classes of fact are known not to survive as structured events, because the old records never held them:

• turn_state rows, which the backfiller has never converted. A gap here has to be closed before any write removal.
• system_note rows, including the eight kinds that are part of the user-visible transcript.
• Provider-native tool activity whose opaque provider output was not retained.

Tool calls, tool results, token usage and permission decisions do convert faithfully; an earlier claim of mine in #4791 that they would be lost was wrong and is corrected there.

The upgrade window is proportional to total history. It will be measured, bounded and resumable after a crash before this ships.

What I am asking for

Objections to the one-way migration itself. Design feedback, file-level review and alternatives belong in #4791, where three of us have already converged on the plan.

I am not holding the work for a comment window. The PR goes up as a draft shortly and review continues there; raise a blocking objection here or on the PR at any point before it merges.