skills: add list-hygiene-judge graph runner

[charlie-morrison]

Adds a list-hygiene-judge graph-runner skill under skills/list-hygiene-judge/.

List hygiene is the judgment between engagement decay and suppression, and the dangerous part is the durable consent-state transition. The skill reads a contact through data-store keyed by the contact as the domain entity, decides verify / suppress / re-permission, and records the transition as exactly one append_event on that contact's stream under idempotency_key + expected_version compare-and-set. It is a thin act{form: review} over content-keyed memory: no operational_proposal envelope, no minted grant.

The skill never sends. send-as is the downstream enforcer: a separate governed run, dispatched by naming, that reads the recorded state at send time and gates delivery.

Judgment order

1. Evidence gate — a missing opens_count / clicks_count / hard_bounces / recency_days counter is a stop, not a zero; an unreadable projection is a stop; a missing bounce_policy is a stop.
2. Ambiguous bounce recovery — hard-bounced and engaged since, inside the decay window. Suppressing drops a live human, re-permissioning ignores a real delivery failure, so neither is taken and it escalates.
3. Suppress — only on hard_bounces > 0 read from the store. The judgment refuses to suppress without bounce evidence.
4. Active unsubscribe marker — refuses to re-permission over it, escalates instead.
5. Re-permission — recency_days past decay_threshold_days, no marker, no bounces.
6. Otherwise no_change, no append.

The append and the read-back are both guarded on decision.writes == true, so every stop path provably emits no event.

Version gate

A stream sitting exactly one event ahead of what the caller read, whose newest event is the very transition this judgment would record, is a retry of our own write, not stale state — the store's append is keyed by idempotency_key and returns the recorded version instead of double-applying. Any other version drift is genuine staleness and refuses to write. The appended event is deliberately free of the prior-projection snapshot so its digest is byte-identical across attempts; without that, a retry lands as an idempotency conflict rather than a replay.

Harness

runx harness ./skills/list-hygiene-judge — three inline cases, green, and green on repeat runs against a warm store:

• sealed_decay_re_permission — 400 days idle over a 180-day threshold, no marker, no bounces → re_permission, one append.
• sealed_hard_bounce_suppress — hard_bounces: 2 → suppress, one append.
• stop_missing_or_stale_evidence — engagement_history absent → stop, both guards block.

Captured run in fixtures/harness-evidence.json.

Note on placement

CONTRIBUTING.md points community skills at standalone packages and reserves this repo for the first-party lane. This PR is opened against runxhq/runx because the bounty that commissioned the skill requires the package files to land here as a public PR; happy to move it to a standalone package instead if you would rather keep the lane clean.