perf: add/delete churn on a small object is ~33x node — shape identity turns over on EVERY op (and the bench's own header now argues the opposite of reality)

[proggeramlug]

The number

bench_dynamic_property_keys.ts's delete-heavy loop: perry 981 ms vs node 30 ms (Mac mini, min-of-7, current main with #9038 default-on). Cross-engine: scriptc 29 ms (node parity), porffor fails to compile it. Sibling issue for the large-object case: #9064 — same root cause, different scale and different dominant term.

Self-contained repro

function deleteHeavy(n: number): number {
  const o: Record<string, number> = {};
  let s = 0;
  for (let i = 0; i < n; i++) {
    const k = "k" + (i % 500);
    o[k] = i;
    s += o[k];
    delete o[k]; // the object never holds more than one live key
  }
  return s;
}
// N = 300_000 → perry ~981 ms, node ~30 ms.

Unlike #9064's stable 500-key receiver, this object oscillates around one live key. Per iteration: an append (new shape id), a read (IC primed against it), a delete (successor id minted, every cache retired). ~3.3 µs/op vs node's ~0.1 µs — the shape-identity machinery IS the loop.

What tombstones already do here (and why it's not enough)

The tombstone threshold (key_count >= 16 before squeezing) means the array grows to 16 slots, squeezes back to ~1, repeats — amortized ~1 slot-copy per delete. Fine. What remains is that all three ops per iteration change or chase shape identity:

• the append publishes a successor id (lineage mint + facts-index insert),
• the delete publishes another one + sweeps stale ids for the owned address,
• and both retirements force the read's and write's ICs to miss every single iteration (js_put_value_set_dyn_ic_miss and the get_field_by_name tail dominate the old profile of this binary, and the descriptor mint/lookup pair — shape_descriptor_by_id + ensure_with_holes — runs multiple times per op).

What to optimize, in order

1. Stable-token deletes + per-key IC validation — perf: populated delete is ~14x node after tombstones — the residual is IC retirement per delete, not the delete itself #9064's option (1). For THIS bench it removes the delete-side retirement; the read IC then survives the delete and only the append still moves identity.
2. Append-side identity churn on re-add: a re-added key appends and mints an id per append. For churn shapes, a small transition cache keyed (predecessor_id, key) → successor id already exists for first-time builds — verify it HITS on the delete→re-add cycle (the tombstone delete's fresh generation may make every re-add transition a cold miss forever). If each k(i%500) cycle re-mints instead of reusing 500 cached transitions, that's the memory-and-time leak to close.
3. Small-object fast path: below the 16-key threshold the per-delete publish could collapse to hole-write + epoch bump with NO descriptor machinery at all once (1) lands — a one-live-key object needs none of it.

Housekeeping in the same change

The bench's header comment (dated 2026-08-28) now argues the inverse of reality and should be rewritten when this is touched: it says the delete penalty is 1.1× ('LOWER than node's — dictionary mode refuted') and overwrite is the real 60× gap. After the write campaigns, overwrite is 13 ms and beats node (19 ms); the delete penalty is 75×. The ratio that used to refute the dictionary-mode argument is now the strongest evidence for #9064's per-key-invalidation work.

Acceptance

• delete_heavy_ms ≤ 3× node on the mini; overwrite_ms stays ≤ node (currently 13 vs 19 — must not regress).
• delete_penalty back under ~2× (node's is 1.6×).
• feat(runtime): O(1) object deletes via tombstones (flag-gated; populated delete 6.5x on, -11% off) #9029 differential battery byte-identical, both tombstone flag states.