node:http/https behavioral parity tail (26.2% → ...) — successor to #2132

[proggeramlug]

node:http / node:https behavioral parity tail

Successor to #2132 (which covered the byte-level diff scope — now closed). This tracks the behavioral gap that re-measurement against Node's own test/parallel corpus surfaced: tests that compile and run but crash, hang, or fail an assertion.

Measurement (Node v22 corpus, scripts/node_core_subset.py --api http https)

	original (#2132 open)	after #2132 fixes
parity	11.6%	26.2%
pass	38	86
runtime-fail	285	240
compile-fail	4	0
diff	2	2

The #2132 fixes (#4903/#4904/#4905/#4906/#4907/#4908/#4909/#4910) cleared the clean single-root-cause clusters (server.address, constructors, missing methods, validation, codegen). What remains is the hard tail.

Remaining 240 runtime-fails by cluster

cluster	count	notes
no-output / server hang	67	framing / keep-alive / timeout hangs; improved by #4964 but still the biggest mechanical bucket
strict-equal mismatch	63	behavioral semantics — largest, grows as upstream fixes unblock tests
other one-offs	56	mostly hang-with-partial-stdout + the distinct bugs below
read-undefined-prop	25	scattered
not-a-function	13	residual lifecycle methods
https TLS cert	6	residual from #4906
missing-exception	5	residual from #4907
ReferenceError not-defined	5	missing globals (MessageEvent, OutgoingMessage, …)

Sub-tickets (distinct root causes split out)

• regression: empty/falsy ClientRequest method '' throws instead of defaulting to GET (over-validation from #4907) #4970 — regression: empty/falsy ClientRequest method '' throws instead of defaulting to GET (over-validation from node:http: missing argument/header/URL validation (assert.throws → Missing expected exception) #4907)
• node:https: js_native_call_method null-pointer crash in closeIdleConnections sweep (test-https-server-close-idle) #4971 — js_native_call_method null-ptr crash in closeIdleConnections sweep
• codegen: derived class over native base — 'Must call super constructor before accessing this' (test-http-client-readable) #4972 — codegen: derived class over native base, super() this-TDZ
• node:http upgrade: 'Object prototype may only be an Object or null' (test-http-upgrade-server) #4973 — http upgrade: Object prototype may only be an Object or null
• node:https: 'tls config unavailable; refusing to listen' under async-dispose (test-https-server-async-dispose) #4974 — https tls config unavailable under async-dispose

Not yet split (need triage)

• strict-equal (63) — behavioral; e.g. test-http-1.0, test-http-agent-scheduling, test-http-agent-keep-alive-timeout-buffer, test-http-blank-header, test-http-chunk-extensions-limit, test-http-client-headers-array. No single root cause — needs per-test diff of actual vs expected.
• no-output (67) — server hangs; needs stderr instrumentation per test (sub-themes: agent lifecycle, outgoing-message framing, keep-alive).

Method

Reproduce locally: vendor the corpus (git clone --no-checkout --depth 1 --branch v22.x --filter=blob:none https://github.com/nodejs/node <dir> + git sparse-checkout set test/parallel test/common test/fixtures), then scripts/node_core_subset.py --root <dir> --api http https --sample-cap 500. Re-measure after each batch to catch regressions (this run already caught #4970).

[proggeramlug]

Re-measured after sub-tickets #4970–#4974 landed

Parity 26.2% → 28.0% (pass 86 → 92, runtime-fail 240 → 234, compile-fail 0). The #4970 regression is confirmed fixed (test-http-client-defaults back to pass).

cluster	26.2%	28.0%	Δ
no-output / hang	67	69	+2
strict-equal	63	67	+4
other one-offs	56	51	−5
read-undef-prop	25	24	−1
not-a-function	13	11	−2
https TLS cert	6	5	−1
missing-exception	5	5	0
ref-not-defined	5	2	−3

The 5 fixes cleared the distinct one-offs (−5 other, −3 ref-not-defined, etc.), but the moving-frontier dynamic persists: no-output (+2) and strict-equal (+4) keep absorbing newly-unblocked tests. Those two buckets (now 69 + 67 = 136 of 234) remain the wall — per-test behavioral/hang work, no shared root cause.

⚠️ Possible new regression to verify

test-http-upgrade-reconsume-stream newly fails ((no output) hang) — it's in the upgrade path that #4973 reworked. Could be fallout from the upgrade-pattern change. Needs a direct run to confirm whether it's a real regression or a pre-existing hang that just surfaced.

[proggeramlug]

Two diagnosed tickets from this session (both with confirmed minimal repros run under Perry):

• regression: http 'upgrade' + new tls.TLSSocket(socket) + destroy hangs the event loop (#5001 fallout) #5010 — regression: http 'upgrade' + new tls.TLSSocket(socket) + destroy hangs the event loop (test-http-upgrade-reconsume-stream, was passing at a34d06212, broke at 50c1a6469 → fix(http): node:http 'upgrade' inherits-pattern + raw-socket upgrades (#4973) #5001/node:http upgrade: 'Object prototype may only be an Object or null' (test-http-upgrade-server) #4973 fallout). Perry run = 15s timeout vs Node exit 0.
• node:http/https: server.ref()/unref() return a raw number instead of this (breaks createServer().unref() chaining) #5011 — server.ref()/unref() return a raw number instead of this, breaking createServer(cb).unref().listen(...) (test-http-request-method-delete-payload, (number).listen). Root cause localized: no ref/unref entry in http_server.rs native table → falls through to generic handle handler; same bug listen had, fixed via node:http: 57 method-surface gaps (Agent/ClientRequest/IncomingMessage/Server.listen) #2129 (ret: NR_PTR).

Remaining (number).<method> cases are distinct (Agent.createConnection returns number; https ALPNProtocols not a Buffer) — noted in #5011, not yet ticketed.