perf: avoid quadratic removed-key scan and skip scalar auto-activation

[deggertsen]

Summary

Reduce work when replacing observable objects and activating their contents while preserving child notifications and lazy activation.

• Skip the removed-key membership scan for keys still at the same position. On the first mismatch, lazily build a Set when both key lists exceed 32 entries, avoiding quadratic scans for large objects. Smaller lists retain includes().
• Bound the same-position lookup by the new key count so an out-of-bounds read cannot hide removal of an undefined Map key.
• Skip primitive values before checking for observables or linked functions during recursive activation.

Set and includes() membership both use SameValueZero semantics, preserving object-key identity, NaN, and signed-zero equivalence. Deletion handling retains its original order.

Regression coverage

• Object replacement and reordering at 32 and 33 keys, plus replacement of 2,000 keys: removed descendants, notification payloads and order, and retained listeners.

• Map object-key identity, NaN, signed zero, and removal of an undefined key after 0, 1, or 33 retained keys.

• Mixed-scalar activation with nested observables, arrays, automatic links, lazy links, and normal functions.

• Replaced the single-sample 5 ms synced-activation assertion with a deterministic test in tests/synced.test.ts. It verifies lazy get/subscribe calls, delivery of 10,000 records without child enumeration, and subsequent access to a nested value. Disabling the plain-data hint in a temporary checkout makes the test fail with 20,000 child enumerations, confirming that it catches eager traversal. The original timing assertion failed on CI at 5.64 ms.

• Stabilized the CRUD last-sync persistence test by awaiting the observable lastSync signal, which is published after metadata is written, instead of assuming persistence finishes within 10 ms. All existing assertions remain.

Validation

• GitHub CI on 7c68964f: Test, Lint, and Format passed (run).
• Full local Jest suite: 26 suites, 1,092 tests passed (npm test -- --runInBand --watchman=false).
• Focused Jest suite: 152 tests passed across CRUD, synced, and performance tests. Bun also passed the synced suite and the metadata regression case.
• npm run typecheck, npm run build, npm run lint:check, npm run format:check, and git diff --check passed.
• The lazy-traversal test was verified against a deliberate regression in a temporary checkout, then that checkout was restored.

Local performance check

Measured September 10 on Apple M4 Pro, Node 26.3.1, using production-mode bundles of base commit 23b5ddeb and the runtime changes in this PR. Object replacement alternates two objects after 50 warmup updates, with 50 measured samples. Per-update milliseconds for 3,000 keys:

Scenario	Base p50 / p90	Updated p50 / p90
Same key order, changed values	3.19 / 3.55	0.50 / 0.58
Reordered keys, changed values	3.21 / 3.67	0.62 / 0.80
All keys replaced	7.68 / 8.24	0.79 / 1.14

First root read of 3,000 string properties measured 0.165 ms → 0.090 ms median (50 warmups, 100 samples). Small-object cases at 16 and 33 keys remained in the low single-digit microseconds, with mixed small differences. These local measurements are not cross-engine guarantees or CI timing assertions.