[boydwold]

Tools outside Ringer that read ~/.ringer/runs/<run_id>.json have two gaps today:

• They cannot tell which version of the run state format they are reading. A reader that misreads a changed field could show a run as passed when it failed, which breaks "displayed data must be true".
• They cannot link a work order back to where it came from. I run Ringer under an orchestrator that plans work in OpenSpec, and a local review app (Referee) reads Ringer's run state to show which task and scenario each work order proves. Run names and task keys cannot carry that link.

This PR adds two small, opt-in parts of one contract for readers of the run state:

• state_version in the run state (now 1). A reader can refuse a version it does not know. The constant's comment says to bump it whenever a field that readers use changes meaning or shape.
• meta in the manifest, at run level and per task: an optional JSON object that Ringer copies into the run state untouched (meta on the run and on each task). Ringer never reads it. Each is capped at 16 KB, so a manifest cannot bloat the state file. A non-object or an oversized meta is a manifest error with a clear message.

Backward compatible: meta is optional and shows as null when absent, and manifests without it run as before.

Proof: tests/test_state_meta.py (13 tests) covers parsing, the size cap, the copy, Manifest.from_path, the snapshot, and the written file. The full suite passes locally on Linux, except test_hud_single_tab.test_dead_then_alive_opens_once_after_spawn, which also fails on current main without this change. The README manifest table documents both fields, and the Contributors section has the credit line that test_contributors requires.

Happy to split this into two PRs (state_version, then meta) if you prefer.

[boydwold]

I added a fix that a review found. ringer.py run --max-parallel N lost the run-level meta, because with_max_parallel rebuilt the manifest field by field and left meta out. It now uses dataclasses.replace, which keeps every field.

The new test test_max_parallel_keeps_run_meta fails before the fix and passes after it. The rest of the suite is unchanged. The only failure is the existing test_hud_single_tab one, which also fails on main.