ORM: no per-transaction memo lifetime — FK probes, CLEARING probes, and relation targets re-read once per row #406

Summary

The ORM has four memo lifetimes. The one several read-amplification defects need — per transaction — does not exist, so three separate call sites re-issue an identical read once per row instead of once per statement.

This is an enabler issue. On its own it changes no behaviour; it exists so the three consumers below stop being three independent hacks.

Current memo lifetimes

The anchor already exists

orm/lifecycle.ts:23:

const ORMLIFECYCLE_INNER_DB = Symbol.for('kitcn:OrmLifecycleInnerDB');

read back at :66-69:

export const getOrmLifecycleInnerDb = (
  ...
  const inner = (db as any)[ORMLIFECYCLE_INNER_DB];

and installed at :675:

Object.defineProperty(wrappedDb, ORMLIFECYCLE_INNER_DB, {

That handle is a stable per-transaction identity — exactly the key a WeakMap<innerDb, Cache> needs.

Consumers this unblocks

1. insert() foreign-key re-probe — one db.get(foreignId) per inserted row for the same parent. update() already memoizes this per statement (orm/update.ts:718); insert() never got the equivalent. Measured 8 reads for 8 rows sharing one FK, where 1 suffices.
2. assertAggregateIndexesWritable (orm/aggregate-index/runtime.ts:3088) — re-probes the by_table_status CLEARING range once per written row. Measured exactly 1:1 with row count (12 rows → 12 probes, 40 → 40), against an empty range every time.
3. Relation-where target resolution — the same target document is re-resolved per parent rather than once per page. Measured 1050 db.get for 2 distinct targets.

The trap

prependWriteBarrier's closure (orm/lifecycle.ts:754) is constructed inside createOrmDbLifecycle, which createOrm calls at module scope (orm/create-orm.ts:301). A naive let probed = false there has isolate lifetime, not transaction lifetime, and would permanently disable the write barrier for the whole isolate — a correctness regression, not just a missed optimization.

The memo must key on the stable per-transaction writer reached via getOrmLifecycleInnerDb, never on the closure.

Suggested fix

Introduce one per-transaction cache owner keyed on the inner writer:

const perTxn = new WeakMap<object, TxnCache>();

with a small typed accessor so the three consumers share one mechanism rather than each inventing a local Map. Land the primitive with one consumer to prove it, then port the other two.

Staleness note

Consumer 2 is safe to memoize: every COUNT_STATUS_CLEARING write happens inside backfill mutations (orm/aggregate-index/backfill.ts) that write only aggregate_* tables, so a user write transaction cannot flip a state row to CLEARING mid-loop.

Environment

• kitcn 0.25.10
• convex 1.44.0
• verified against origin/main at 00f7f121