Add a durable orchestration run manifest and deterministic resume command

[tomvaillant]

Problem

Splash orchestration state is spread across the conversation, accepted.json, report.json, generated artifacts, and implicit knowledge of which command ran last. After a restart, context compaction, terminal failure, or handoff to another journalist, there is no deterministic way to answer:

• Which gate is this run at?
• Which elements are accepted, produced, reviewed, approved, or delivered?
• What is the next valid action?
• Is the current report/artifact still the one that was reviewed?

The agent currently reconstructs this from prose and files, which is fragile and not auditable.

@rmdms

Suggested solution

Add a versioned run manifest and a deterministic resume command, for example:

bun skills/splash/scripts/resume.mjs exports/<slug>/run.json

The manifest should be the canonical orchestration ledger and contain no secrets. It should record:

• Run ID, schema version, project/input paths, and hashes of article/data inputs.
• Confirmed CADRAGE answers and per-element confirmed takeaways.
• Accepted proposal IDs and hashes.
• Per-element gate state: proposed, accepted, produced, reviewed, previewed, approved, delivered, blocked, or dropped.
• Current artifact/report hashes, generation timestamps, and provenance links.
• Structured review findings and overrides.
• Requested deliverables and completed delivery artifacts.
• Current gate plus the finite set of valid next actions.

Every gate command should update the manifest atomically after success. Failed commands should append a bounded failure event without advancing state. accepted.json and report.json can remain pipeline artifacts, but the run manifest owns orchestration progress and points to their exact versions.

resume should validate hashes/provenance, print a concise journalist-facing status, and emit the exact next command(s). It must never infer completion from conversation text alone.

Acceptance criteria

• Closing and reopening a run resumes at the same gate with the same valid next actions.
• Multi-element runs track each element independently without dropping unfinished work.
• A newer artifact, stale report, or changed accepted spec is detected and cannot inherit prior review/approval.
• Manifest updates are atomic and recover safely after interruption.
• No API keys, tokens, article secrets, or full credentials enter the manifest.
• Tests cover every gate transition, invalid transitions, crash recovery, stale provenance, schema migration, and multi-element resume.