feat(runtime): node:stream + node:stream/web (Readable/Writable/Transform + WHATWG) (#1545)

[proggeramlug]

node:stream + node:stream/web parity (#1545)

node:stream and node:stream/web were floored/skip-listed in the #812 matrix. This makes the in-repo test-parity/node-suite/stream suite (801 differential tests vs node v26.3.0) go 739 → 761 PASS (+22) with zero regressions.

What now works (root causes)

1. No more spurious Uncaught (in promise) on WHATWG stream abort/error/cancel (+13).
WritableStream.abort(), ReadableStreamDefaultController.error(), reader/writer cancel, pipeTo to a failing sink, and TransformStream cancel-propagation all reject the stream's internal closed/ready control promises. Those had no user-attached reaction, so Perry reported them as unhandled — Node marks them markPromiseAsHandled at creation. Added js_promise_mark_internally_handled (a persistent, hot-path-free set consulted only when non-empty) and applied it to every reader/writer/stream closed + ready promise at creation.

2. Async stream operations that observe a rejection by polling no longer leak it (+6).
Readable iterator-helpers (map/filter/every/some/…), pipeline(), and compose() drive an async callback by polling the returned promise's state and reading reason on rejection — without attaching on_rejected. They now mark_rejection_handled the promise they consumed, matching Node (the user still sees the error through the normal channel; only the duplicate internal leak is silenced).

3. Readable.read() (no size) returns the head chunk, not the whole buffer (+4).
Matches Node's howMuchToRead(NaN): a readable-event drain loop (while ((c = r.read()) !== null)), for await, unshift()-then-read(), and Readable.toWeb() now preserve chunk boundaries instead of getting one concatenated blob. Sized read(n) still spans chunks.

4. Async data-listener rejections are swallowed (+1).
Node's Readable neither captures an async data listener's rejection to error nor reports it as unhandled; Perry now marks it handled to match.

node-suite stream: 739 → 761 PASS / 0 CFAIL

The remaining 40 "diffs" are not Perry failures: 33 are tests where node itself exits non-zero (intentional TransformStream HWM-0 backpressure deadlocks → unsettled top-level await exit 13, or Readable.from(promise)-style ERR_INVALID_ARG_TYPE throws), which gen_feature_matrix.py scores as NODE-FAIL and excludes. The genuine remainder (subclassing class X extends Readable, an extra readable event in one objectMode case, a repipe-after-unpipe double-delivery) are left as follow-ups — each is a distinct, higher-risk area.

Differential repros (node v26 vs Perry — byte-identical)

r1_web_cancel.ts      first: chunk-1 / cancelled cleanly        (no Uncaught on cancel)
r2_pipeline_error.ts  pipeline err: boom                        (async generator throw → cb)
r3_chunk_boundaries.ts chunks: 3 => ["header\n","body-line-1\n","body-line-2\n"]

Regression guard

• test-parity/node-suite/stream: 0 regressions (every previously-passing test still passes; pass-set diff is empty).
• test262 built-ins language shard 0/16 (scripts/test262_subset.py, --jobs 40): pass=1873, diff=14, runtime-fail=92, compile-fail=5, parity 94.4% — byte-identical before and after (the only core change, the unhandled-rejection set, is empty for any program that never touches streams, so non-stream behavior is provably unchanged).

Files

perry-runtime: promise/then.rs (+js_promise_mark_internally_handled), promise/mod.rs, lib.rs, node_stream_iter_helpers.rs, node_stream_pipeline.rs, node_stream_readable_read.rs, node_stream_event_emitter.rs. perry-stdlib: streams.rs (+internal_promise() helper), streams/{byob,transform,writable}.rs. (+134 / −48)