Make lineage caches family-aware and atomic

[iam-brain]

Question

Implement lineage-aware incremental caching so parsed physical documents remain reusable while reconciled family results are invalidated whenever membership or member content changes, active-to-archive moves do not recount unchanged documents, schema changes cannot mix legacy and lineage totals, and cancellation never publishes partial cache or report state.

Locked constraints

• Cache parsed documents independently from reconciled accounting.
• Never treat a per-file accounting result as authoritative lineage output.
• Key reconciled results by a deterministic family fingerprint covering normalized Codex-home scope, member identities, and document content signatures.
• Recompute only affected families when possible; discovering a missing parent invalidates that family.
• Treat sessions and sibling archived_sessions under one normalized Codex home as one source; never merge separate Codex homes.
• Cache updates and published totals are atomic after successful completion.
• Accounting mode and schema version prevent incompatible cache reuse.