[wfaithfull]

This is M10 step 2 (DESIGN.md §5.7). It makes the id map generic over how it stores ids, with no change in behaviour, so step 3 can add TextKeys without touching callers again.

Shape

• IdMap<K: IdKeys> holds keys: K and the tombstones. The rebuild's merge walk (tombstone, keep or append) stays in IdMap::merge, so a new key storage only has to supply its storage.
• trait IdKeys:
  • associated types: Key: ?Sized + Ord, Ids: NodeIds, Rebuild: KeysRebuild<Self>;
  • lookups in both directions: dense_id and key, with external_id returning a reference;
  • in_order, renumber, compacted, and a rebuild that reserves exactly.
• FixedKeys<E> holds today's pairs and inverse unchanged. The parts that only apply to fixed ids stay on IdMap<FixedKeys<E>>: segments, shared views and from_parts.
• Now generic over K: EdgeIntake, GraphVersion, compaction, diff, the degree counts and distance alignment.
• pg_warren only changes to IdMap<FixedKeys<E>>. Splitting IdDatum is left to step 4.
• Small changes:
  • IdMap::renumber builds the new tombstones first, so a failed memory charge changes nothing.
  • GraphVersion gets hand-written Clone/Copy, because a derive would require K: Copy.
  • The fuzz target decode_id_map is updated.

§5.7 records the trait as built.

Checked locally

• warren-core: 651 unit tests and 5 integration tests pass, the same counts as before.
• Miri, with CI's flags and the pinned nightly: passed.
• fmt, both clippy runs with -D warnings, rustdoc with -D warnings, and the core-isolation check.
• cargo pgrx test pg16: 636 passed, before main (with set_owner) was merged in. CI covers the merged tree.

---

Generated by Claude Code