fail_closed on an UNKNOWN leaf does not fire for non-blocking (warn) rules

[Fszta]

Summary

Under on_missing_meta: fail_closed, an UNKNOWN (unresolvable) leaf fires the rule only when the rule is blocking (_is_blocking, parrant/lineage/policy.py ~L924-950). A rule whose action is warn (non-blocking) therefore silently allows on an UNKNOWN leaf — it does not warn. So warn + fail_closed does not actually fail closed.

Why it matters

It's a footgun for compliance policies authored as advisory (warn-only) during rollout: an author writing on_missing_meta: fail_closed reasonably expects "warn me when you can't prove it's safe," but gets silence. A real PII-exposure policy hit exactly this (see #123): once lineage was unresolvable, the fail-closed PII rule returned ALLOW instead of surfacing anything.

Options to consider

1. Uniform severity: an UNKNOWN under fail_closed fires the rule at its declared action severity (warn→warn, block→block) — most intuitive reading of "fail closed", but changes behavior for existing warn rules (possible noise in meta-poor repos).
2. Explicit 'unproven' surface: instead of firing, add an explicit unproven/indeterminate count + list to the report (and JSON verdict) so unresolvable leaves are visible without warning-flood. Visibility without coupling to the fire path.
3. Status quo + docs: keep block-only firing; document that warn rules do not fail closed on UNKNOWN, and that fail-closed enforcement engages at block-promotion.

Leaning toward (2) as the default with clear docs; (1) behind a knob if wanted.

Repro

The deferred-build fixture + runners from the #123 investigation reproduce an UNKNOWN leaf under a warn+fail_closed rule returning ALLOW.