fix(runtime): retain possible-effects observations on failure

[danielkov]

Summary

Retain bounded possible-effects observations across child ACP execution and live root sessions, including detached fork abandonment. Store conservative snapshots in version-4 fatal diagnostics while preserving native tool outcome classification.

Motivation

Related to #48. A failed or cancelled prompt can already have emitted output or invoked tools; losing those observations hides uncertainty from diagnosis.

Impact

This remains a draft, preparatory implementation. Parent-visible typed failure, cancellation, and background metadata still require compatible upstream transport releases. These facts do not establish replay safety and introduce no automatic session recovery.

Technical details

Local observations are cumulative within one live session owner, not exclusive to the failing prompt or reconstructed from persisted history. Local execution receipts reflect actual invocation entry and terminal return; ACP lifecycle statuses remain reports. False means only not observed, and failure snapshots always declare incomplete observation. The bounded metadata excludes prompts, assistant text, tool arguments/results, and provider payloads.

Startup and prompt-preparation failures retain their recording owner without duplicate receipts. Cleanup or flush errors that replace a driver error receive a separate finalization diagnostic. An acknowledged fork handoff assigns recording responsibility to either the caller or detached owner, retaining observations if delivery is abandoned.

Readers retain legacy missing-effects defaults and version-3 effects compatibility while rejecting malformed fields, unknown sources, and false completeness claims. Parent-side observation records are distinct diagnostics, not reconstructed child fatal identities.

[kit-code-agent]

Abandoned fork calls can discard captured possible-effects observations before they reach the local fatal diagnostic.