feat(#579): org security postures + provenance-labelled inbound scree…

[sneumannb5]

What

Org security postures (#579): two-axis model (dangerous/auto/strict) with tighten-only org floor + provenance-labelled inbound screening (LLM judge or HTTP proxy) before the turn runs.

Why

Adds the ingress mirror to Privacy Shield: injection-screening tier + org-wide posture scopes can only tighten. Quarantine on enforce, fail-open-with-evidence when unscreenable, shadow mode to observe.

Test plan

• typecheck @omadia/channel-sdk + @omadia/orchestrator — clean
• securityPosture*.test.ts — 47/47; verifier suites — 126/126
• core-decoupling ratchet held at 3294
• full middleware suite after main merge

Risk / blast radius

• New setup fields security_posture{,scope}, security_screen{mode,url} (defaults auto/enforce). No schema/migration.
• Gate at every turn ingress; verifier re-runs exempt (screen once per user turn).
• Limitation: attachment screening is descriptor-only until mid-turn body extraction lands — auto reachable but weak.

Naming-decisions still pending

• Tool-approvals axis deferred — strict falls back to auto-screening; flip when approvals ship.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}