fix(bare): recover after a run stops before provider start #343

What Bare threads can refuse later messages after a run is stopped or fails before its provider turn starts.
Why Missed-history recovery includes T3 error and interrupt-result records, but the bare prompt boundary rejects every missed item that is not a user message.
So what Bare recovery needs a separate fix and focused regression tests; #285 will retain and link this known limit rather than redesign recovery.

Outcome

A bare thread accepts the next message after a pre-start Stop or session-open failure, without injecting T3 status text, losing undelivered user text, or repeating an earlier delivered attempt. This supports the Objective, making Promachos and Spectrum work reliably in Chromeria while preserving bare prompt isolation.

Evidence and verification limit

Found while integrating #285 / PR #300 with #289 / PR #335. This report is source-traced at main 2ba387388074ac4ee09cc256f9800cdd9ce2b520; no native reproduction is claimed. The mechanism exists on main without #285. Searches of open bare issues and all issues for the exact error, bare recovery and bare interrupted found no separate matching open issue.

• ProviderTurnStartService selects failed/interrupted earlier runs whose active attempt has no provider turn, then reads their turn-start history.
• ProjectionStore.getTurnStartHistory includes error and run_interrupt_result; historicalMessage retains both.
• Before-start Stop writes run_interrupt_result with Run interrupted before provider start. Session-open failure writes an error record.
• The bare delivery boundary rejects those records with bare history transfer unsupported: missed history is not user text. The failed retry remains eligible as missed history for the following message.
• Restore explicit urgent interruption after Claude steer queuing #285 adds triggers in the replacement window and its truthful Message not delivered failure, but the original pre-start Stop and open-failure paths already exist.

Reproduction to establish in the fix: send to an isolated bare thread, stop before provider start, then send another message; also cover session-open failure. Never use live userdata.

Acceptance criteria

• A persisted-event integration test reaches pre-start Stop through the real command/effect path and proves the next turn delivers the missed user text and new text with no T3 status prose.
• A missed two-attempt run resends only text belonging to its current undelivered attempt, preserving the earlier delivered attempt without duplicating it.
• Named additions preserve their existing native-session carrier semantics.
• Genuine unsupported non-user history remains a specific failure unless its semantics are explicitly designed.
• Tests await persisted events or Deferred milestones, with no sleeps.

Non-goals and blockers

No change to #285 cancellation scope, requester notes, bare provider support, launch controls or receipt policy. No implementation is included in this report. The precise recovery semantics and a failing transition test remain to be designed; no external blocker is known.

Elon record

Requirements and who asked: User #289 requires prompt isolation and supported bare providers; #285 requires urgent interruption and truthful failure. AGENTS.md Work requires filing adjacent owner defects.
Deleted: A speculative recovery patch inside #285, silent dropping of all history, and a claim of native reproduction without one.
Bottleneck: The bare missed-history boundary treats internal status records as model history; its owner must decide which records to omit and which attempt's user text to retry.
Checked myself: Read main's missed-run selection, final bare delivery guard, ProjectionStore history query, historicalMessage conversion, and before-start interrupt producer; checked matching open/all issue searches.