[suprie]

Recover interrupted orchestration runs and edit budgets

Summary

Continue unfinished queued loops after interruption without discarding existing work. Store run plans and execution snapshots in per-user SQLite shared by linked worktrees, and let users raise token or time limits before continuing.

What changed

• Import validated legacy run-plan/execution YAML once into a schema-versioned SQLite store outside the repository. Resolve linked worktrees to the main checkout and enforce execution revision checks in database transactions.
• Preserve managed branches, worktree edits, completed phases, attempt counts, provider session checkpoints, and failure diagnostics during recovery. Resume compatible Claude/Codex sessions; fall back only for explicit missing-session cases.
• Add inactive-run budget editing through typed IPC and a run queue form. Reject zero limits and preserve saved progress.
• Show Codex sub-agent call results and child status separately, without treating a completed delegation call as a completed parent turn.
• Apply shared ASD-STE100 writing guidance alongside optional role charters for both providers.
• Update role-based orchestration PRD completion bookkeeping, compact project memory, and document operational recovery and migration limits.

PRD

• docs/epics/role-based-orchestration/prd-role-foundations.md
• docs/epics/role-based-orchestration/prd-agent-handoff.md
• docs/epics/role-based-orchestration/prd-parallel-scheduling.md
• docs/epics/execution-foundations/operational-recovery.md

This change builds on the merged role foundations and handoff work. The parallel-scheduling PRD status update is bookkeeping for its separate implementation in PR #113; this PR does not introduce that scheduler.

Decisions and limits

• Centralize mutable run/execution snapshots; retain authored specifications and decisions in Git.
• Preserve legacy YAML as the original import snapshot; it is not maintained as a current export.
• Project identity follows the canonical main-checkout path. Repository relocation needs a later stable-ID migration.
• Delivery retry records, handoff files, session logs, event replay, and backup/restore UI retain their existing scope or remain follow-up work.
• Local execution/delivery YAML changes are excluded from this PR.

Test plan

Verified after integrating the handoff implementation from merged PR #112:

• cargo test --lib -- --test-threads=1 — 768 passed, 9 ignored.
• cargo fmt --check.
• cargo clippy -- -D warnings.
• npm run build.
• npm run test:frontend — 36 passed.
• Recovery tests use real temporary Git worktrees/databases and a fake Codex transport.
• Manual: interrupt and restart a live Claude/Codex queued run, raise its budget, and confirm continuation preserves edits and completed phases.

Live-provider tests were not run. Rust test compilation reports existing unused-result warnings in agents.rs; production Clippy passes.