fix(otel-thread-ctx): detect AsyncContextFrame by reading CPED natively

[szegedi]

Follow-up to #397, implementing @nsavoire's suggestion from this comment: gather the evidence natively by reading CPED, rather than inferring it from AsyncLocalStorage#run's behaviour.

Of the three shapes I floated in r3776016667, this is the third — cpedMapContains(key, value). Rationale for that choice below, since it wasn't obvious to me at first either.

Why the old probe was worth replacing

#397's detection overrode enterWith on a throwaway AsyncLocalStorage and checked whether run() dispatched through it:

probe.enterWith = () => { delegated = true; };
probe.run(0, () => {});

That run() reaches the instance's enterWith property is unspecified. Anything patching AsyncLocalStorage can break it — dd-trace-js patches async-context machinery itself — and the resulting false negative is precisely the failure #397 existed to fix: ThreadContext#enter() throwing from an inline diagnostic-channel subscriber, surfacing in application code.

What replaces it

const probe = new AsyncLocalStorage();
const sentinel = {};
let bound = false;
probe.run(sentinel, () => { bound = addon.cpedMapContains(probe, sentinel); });
probe.disable();
active = bound;

With ACF active, Node installs an AsyncContextFrame — a JS Map keyed by the storage instance, valued by its store — as the running continuation's CPED. Nothing writes the slot otherwise. I verified the keying empirically rather than assuming it.

This is the same slot and the same "is it a Map" question that WallProfiler::SetContext already asks before storing a context, and the key is the one whose identity hash we publish as otel_thread_ctx_nodejs_v1.als_identity_hash for the out-of-process reader to look up. So a wrong answer here means both consumers are already broken — the detection now depends on exactly what the consumers depend on, and nothing more.

Why key+value rather than cpedIsMap

CPED is a general embedder slot. A Map that some other native addon left there must not be able to answer for us — that would readmit the silent false positive #397 killed, where the writer looks healthy from JS while every reader sees records nothing updates.

This is load-bearing, not theoretical: mutating the helper to a bare IsMap check fails three of the five new tests and none of the pre-existing ones. Without them, the weaker version would pass the whole suite.

Why not getCped

Returning the live frame to JS would hand any caller who can require the addon an object whose mutation corrupts async-context propagation process-wide. It also moves the check into JS, where instanceof Map is realm-sensitive (the frame is a SafeMap subclass), so it'd need to be acf?.get?.(probe) === value to be safe — same answer as this, larger blast radius.

Implementation notes

• Public v8::Map::Get, not map-get.hh. Tempting to reuse the raw OrderedHashMap walk since that's what the reader does, but it would conflate "is ACF on" with "is our layout knowledge still correct", reporting a V8 layout change as ACF being unavailable. Layout has its own coverage.
• In binding.cc, not wall.cc, so it works on Windows. wall.cc's #ifndef _WIN32 block is there for SIGPROF and the v8::base::TimeTicks symbol trick; a CPED read needs neither. No JS fallback anywhere.
• NODE_MAJOR_VERSION >= 22, returning false below — the correct answer there, rather than a missing export the JS side has to reason about.
• Total by construction. No context entered, slot unset or not a Map, key absent, or a malformed call all yield false, never a throw. ensureHook() calls this.

Verification

All six detection routes, on both Node lines:

Node	route	detected
24.18.0	default	active
24.18.0	--no-async-context-frame	inactive
24.18.0	NODE_OPTIONS=--no-async-context-frame	inactive
22.23.2	default	inactive
22.23.2	--experimental-async-context-frame	active
22.23.2	NODE_OPTIONS=--experimental-async-context-frame	active

124 passing on macOS; 175 passing / 2 pending in npm run test:docker (170 before — the five new tests).

Not in this PR

The >=22.7.0 nits from the same review are untouched, so they can be applied independently. Worth noting the direction agrees with them: with the capability observed directly, the detection path carries no version arithmetic at all, and version reasoning survives only in asyncContextFrameHint()'s advisory text — where major < 22 should become 22.7.0.

[github-actions]

Overall package size

Self size: 2.53 MB
Deduped: 3.24 MB
No deduping: 3.24 MB

Dependency sizes

| name | version | self size | total size | |------|---------|-----------|------------| | pprof-format | 2.3.1 | 504.33 kB | 504.33 kB | | source-map | 0.8.0 | 185.66 kB | 185.66 kB | | node-gyp-build | 4.8.4 | 13.86 kB | 13.86 kB |

_{🤖 This report was automatically generated by heaviest-objects-in-the-universe}