Temporal validity intervals on entities + relations (as-of graph queries)

[addadi]

Pitch

memory_atoms already carries validity columns — valid_from_ms, valid_until_ms, last_verified_at_ms (src/migrations.zig:153-155, mirrored in the Postgres schema at :610-612). entities and relations do not: they only have created_at_ms (entities at :135, relations at :236). This means an atom can be time-scoped, but the entity it references and the edges connecting entities cannot. Graph traversal today can only filter on status (proposed/superseded/...), not on "was this edge true at time T".

I'd like to build this if it's wanted upstream, and open a PR. Filing this first to check direction before writing code.

Proposed design (valid-time only, not full bi-temporal)

1. Schema — add nullable valid_from_ms + valid_until_ms to entities and relations, in both the SQLite and Postgres schemas, copying the memory_atoms column pattern exactly. New migration; no edits to shipped ones. Semantics: NULL valid_from = since-creation; NULL valid_until = still valid. Transaction time stays created_at_ms (no second time axis).
2. Query surface — optional as_of_ms parameter on the graph commands (traverse/path/relations), filtering rows to valid_from_ms <= as_of < coalesce(valid_until_ms, inf). Default (no param) = current behavior, zero breakage.
3. Supersedes interplay (the open design question) — today supersedes is expressed as a relation type + a 'superseded' status. If we add validity intervals, should writing a supersedes edge automatically close the target's valid_until_ms = now? My lean is yes (status + validity should agree), but I want a call from the maintainer before encoding it.

Scope / non-goals

• Not proposing a Cypher parser or in-process graph algos — graph_projection (FalkorDB/Neo4j) already covers that per the README.
• Not touching artifacts (it has last_verified_at_ms but a different lifecycle).

Validation plan

• Migration applies cleanly on SQLite + Postgres; existing DBs upgrade without data loss.
• traverse with as_of_ms excludes expired relations; without the param, behavior is unchanged (regression tests).
• `zig build test-local --summary all` green.

Open to feedback on the semantics, the `as_of` naming, and whether `supersedes`-closes-validity should be automatic or explicit. Happy to scope down (e.g. relations-only, skip entities) if that's a better first cut.