session: make checkpoint persistence failures observable and policy-driven

Problem

Main-loop and delegated-session checkpoints are best effort: save errors are discarded. A run can report success while its latest main, Subagent, Parallel, or Team snapshot was not persisted, leaving resume/inspection stale without telling the operator.

Desired outcome

Give persistence failures an explicit, bounded, operator-visible policy without turning transient storage failures into duplicate model/tool execution.

Scope

• Inventory every checkpoint site and classify which saves are required vs best effort.
• Define retry/backoff and terminal behavior for failures before a run, at turn boundaries, while awaiting approval, and at child termination.
• Surface failures through the existing diagnostics/event taxonomy without leaking transcript content or adding noisy duplicate lines.
• Preserve cancellation-detached terminal child saves and bounded shutdown.
• Add tests proving a reported successful durable outcome cannot silently depend on a failed required checkpoint.
• Document degraded behavior for stores that become unavailable mid-run.

Acceptance criteria

• Save failures are never silently discarded.
• The policy avoids replaying an already-started model stream or tool call.
• Main and delegated sessions follow one documented classification of required/best-effort checkpoints.
• Operators can identify the affected session ID and checkpoint without exposing prompts/tool arguments.
• Offline failure-injection tests cover main, foreground child, background child, Parallel, and Team paths.

Relevant code

• engine/agent/loop.go
• engine/agent/subagent.go
• engine/agent/parallel.go
• engine/agent/teamsupervisor.go
• internal/adapter/server/service.go

Start by tracing checkpoint sites in engine/agent/loop.go, subagent.go, parallel.go, teamsupervisor.go, and internal/adapter/server/service.go, then inventory their required or best-effort behavior. Define the bounded failure policy and diagnostics flow without replaying started work; done means failure-injection coverage for main, child, Parallel, and Team paths plus documented degraded behavior.