feat(runtime): persist complete thread continuation state (YEF-1061)

[Yevanchen]

Summary

• Define warm/cold continuation equivalence for Task Agent Threads in the PRD and architecture.
• Commit the complete Cattle Thread workspace through the existing Cloudflare Sandbox Backup path after a successful turn and before lease release.
• Atomically bind the ready checkpoint to its Run and promote the provider-native resume cursor; block follow-up admission and idle recycle until that boundary is ready.
• Grandfather pre-rollout Threads, restoring their recorded artifacts on cold continuation, then enable strict Run-bound checkpoints after their first successful post-rollout turn.
• Return a typed SESSION_RUN_CHECKPOINT_PENDING / Public API readiness_blocked 409 for follow-ups that arrive while checkpointing is incomplete.
• Restore the exact Thread checkpoint on cold activation, exclude attachment mounts and transient credentials, and return actionable missing/corrupt-state errors.

Why

• Cattle continuation previously restored only recorded artifacts and bounded message history, so a recycled Sandbox lost files outside outputs/, Git state, installed workspace dependencies, and provider-local resume state.
• A user's next message should observe equivalent Thread state whether the Sandbox remained warm for ten minutes or was completely recycled for twenty days.
• Existing Threads must remain usable across rollout even though their completed Runs predate Run-bound checkpoint metadata.

Verification

• Commands:
  • CI=1 bun run check (passed: formatting, docs, lint, type checks, all package tests, and Public API contract checks)
  • bun run --filter @mosoo/api test (1240 passed)
  • bun run commit:check (passed)
• Manual steps: N/A; cold restore, 20-day retention, rollout compatibility, failure, retry, admission, idle sweep, and idempotency are exercised with deterministic fixtures.
• Not run: A live production Cloudflare Backup/Restore cycle.

Impact

• User/API/contract changes: Task Agents now promise Thread-scoped workspace and native-resume durability. A follow-up during checkpointing returns a retryable 409 instead of an internal 500; no new public endpoint is added.
• Generated files / GraphQL / DB / lockfile: Extends D1 migration 0011_cattle-terminal-checkpoints.sql with the additive session.workspace_checkpoint_required rollout marker and updates its generated snapshot; no GraphQL or lockfile change.
• Env or config changes: None.
• Risk and rollback: Pre-rollout Threads default to compatibility mode and retain artifact recovery. A new successful Cattle turn enables strict checkpoint enforcement; checkpoint failure intentionally blocks the next turn. Rolling back application behavior leaves additive columns and the partial index inert.

Review

• Closest review areas: rollout-marker transition, terminal event/checkpoint/lease ordering, D1 batch atomicity and concurrent retry cleanup, exact-Thread restore selection, admission/idle guards, and credential-safe checkpoint preparation.
• Known trade-offs: Live processes, sockets, kernels, attachment mounts, and short-lived credentials are intentionally excluded. Background self-healing after terminal checkpoint retry exhaustion is tracked in YEF-1067.

Closes YEF-1061