With no async hook and no executionAsyncResource() user, every InternalCallbackScope pushes and pops the async ids on the stack, only for executionAsyncId() and triggerAsyncId() to read the top. This swaps the two ids in place instead: they are all that those two and process.nextTick() read, so the readers pay nothing.

executionAsyncResource() is the only reader of the stack. The scopes that skip it form a chain in AsyncHooks, counted in a new field kLazyScopes, and JS asks C++ for the innermost one only while that count is not zero. The checks for a corrupted stack stay: test/async-hooks test-improper-order and test-emit-*-on-destroyed fail without them, because the main script runs in a scope that starts before any hook.

compare.js on napi/make_callback, 30 runs, Linux x64, one core, on main (3e5c8af, with #66395):

                                                     confidence improvement accuracy (*)   (**)  (***)
napi/make_callback n=1000000 type='AsyncResource'            **      6.57 %  ±4.56% ±6.07% ±7.91%
napi/make_callback n=1000000 type='Call'                            -2.91 %  ±5.61% ±7.51% ±9.86%
napi/make_callback n=1000000 type='MakeCallback'              *      7.47 %  ±5.64% ±7.50% ±9.77%
napi/make_callback n=10000000 type='AsyncResource'          ***      8.54 %  ±3.42% ±4.55% ±5.93%
napi/make_callback n=10000000 type='Call'                           -1.02 %  ±4.35% ±5.79% ±7.54%
napi/make_callback n=10000000 type='MakeCallback'           ***      7.30 %  ±2.85% ±3.80% ±4.96%

MakeCallback goes from 128.0 to 119.3 ns at n=1e7. test/async-hooks, the async hooks, AsyncLocalStorage, AsyncResource and nextTick tests in test/parallel, and the new addon test pass.

The new field and the chain are the part to look at first, the design was discussed in nodejs/performance#24.

Disclosure: I used Opus 5.5 (Max) as coding assistant