[P1] CLIPS working memory serialization/deserialization for crash recovery

[se-jo-ma]

Problem

FathomRouter.route() (nautilus/core/fathom_router.py:132) rebuilds CLIPS engine state from clear_facts() + asserts on every request. There is no snapshot/restore of working memory. On crash, all in-memory CLIPS state (cumulative exposure facts, relationship facts, multi-request rule state) is lost — only what was flushed to the audit log survives.

Design references

• design-docs/05-ecosystem-roadmap.md:423 — v1 roadmap item: "Working memory serialization/deserialization"
• design-docs/05-ecosystem-roadmap.md:433 — related v1: "Health check endpoint (readiness gating on deserialization)"

Code locations to modify

• nautilus/core/fathom_router.py — add snapshot() returning serialized fact set + restore(snapshot) for replay
• nautilus/core/broker.py::setup() — restore from persistence on startup before serving traffic
• nautilus/core/broker.py — periodic snapshot write (configurable cadence)
• /readyz endpoint should fail-closed until restore completes
• Schema for snapshot storage: extend session_pg.py or separate table

Acceptance

• Snapshot captures complete CLIPS fact set including session exposure multi-slots.
• Restore reconstructs identical fact set; rule evaluation deterministic before vs after.
• /readyz returns 503 during restore.
• Integration test: assert facts, snapshot, restart, restore, assert facts identical.

Priority

P1 — listed as v1 hardening item, unimplemented. Complements WAL (#22) — WAL covers in-flight requests, snapshot covers steady state.