Wave 1: add durable run artifacts and append-only execution history to agent_runtime

[tomanizer]

Parent umbrella: #195 (#195)

Problem

The runtime currently persists useful state in SQLite, but the main operational model is still "latest known state per work item" rather than "durable history of executions with replayable artifacts".

That makes it harder to:

• debug retries
• compare attempt N to attempt N+1
• recover after backend failures
• inspect exactly what context and backend output produced a decision

Why This Belongs In Wave 1

Reliable automation depends on durable artifacts and run lineage. Without that, the runtime remains a convenient dispatcher rather than a trustworthy control plane.

Scope

• define a durable run artifact layout under .agent_runtime/
• persist per-run artifacts such as:
  • rendered handoff bundle
  • backend command metadata
  • stdout/stderr where available
  • parsed outcome payload
  • key PR metadata snapshot where relevant
• move toward append-only execution history keyed by run_id
• keep a current-state projection for fast routing, but do not rely on it as the only history
• add inspection commands or documented inspection queries for run history

Out Of Scope

• automatic merge
• event-driven orchestration framework changes
• full review-comment ingestion if that requires a separate issue

Dependencies

• shared handoff bundle strongly recommended

Mode Impact

Manual

• manual runs can store and replay outcome artifacts instead of relying on shell-quoted summary flags only

Semi-Manual

• runtime-managed runs become easier to inspect, compare, resume, and complete

Autonomous

• automated runs gain the durable execution history required for replayability, observability, and policy-gated resume behavior

Acceptance Criteria

• each runtime execution has a durable artifact directory keyed by run_id
• the runtime persists enough information to reconstruct what was asked, what backend ran, and what came back
• execution history is append-only at the run level even if a current-state view remains for routing
• retry analysis across multiple runs for the same work item is possible without reading overwritten state only
• documentation explains where operators inspect run artifacts

Evidence

• agent_runtime/storage/sqlite.py
• agent_runtime/orchestrator/graph.py
• agent_runtime/orchestrator/parallel_dispatch.py
• agent_runtime/manual_supervisor_workflow.md

Notes For Work-Item Decomposition

Likely split:

1. run artifact directory design
2. append-only execution-history schema changes
3. current-state projection updates
4. inspection tooling and tests