[oxura]

Fixes #1268. Native Chan is now an owned heap row with an owned FIFO of two-Term nodes, using existing ARR reference counting and iterative term_drop. The global row table, generation handles and deferred row recycling are removed. Each effect consumes its handle; parked operations retain it until wake/deadline. Close still wakes waiters and preserves buffered FIFO contents for copied handles. Receiving transfers a value out, and a full-channel dequeue reuses its node when waking a queued sender. Capacity is logical: creating a channel no longer reserves/touches a capacity-sized ring. No new runtime tag, finalizer framework or garbage collector.

Actual Linux/clang 22.1.8 verification:

• 500,000 room-one buffered send/close iterations, --threads 4, exit 0 and done: baseline peak RSS 40,728 KiB; treatment 2,232 KiB. Empty-close control: 2,268 -> 2,140 KiB. Exact commands use /usr/bin/time -f 'RSS_KIB=%M elapsed=%e'. Single-run elapsed samples are not a general performance claim.
• 26 actual native consumer runs at 1/4 threads pass: copied closed-buffer FIFO, boxed values, final parked handle/deadline, rendezvous, try-send/receive, wake/timer cancellation, fork/join and channel-only deadlock termination.
• Earlier clang AddressSanitizer runs at 4 threads reported no diagnostics for ordinary allocations such as waiter timers. ASan cannot see Bend heap row/node bounds or prove their reclamation. This x86 runtime pins registers; those builds use -fomit-frame-pointer -mllvm -x86-use-base-pointer=false with -fsanitize=address.
• Both new consumer fixtures also pass the actual Bun CLI's unchanged JS backend.
• Repository gate: PASS 54 / 54. Protected files and gate limits unchanged.

There is deliberately no permanent RSS threshold: standalone before/after/control measurements prove reclamation; deterministic regressions cover ownership transitions without depending on platform RSS or sanitizer overhead. Native Chan's opaque foreign representation changes, documented in EFFECTS; foreign effects already have no ABI promise. Lifetime follows the existing native reference-counting model, not tracing collection of cycles. No Metal/CUDA run is claimed.

Review follow-up (a781bf7)

Preserved maintainer d95603c: Chan handles are born RFC-wrapped so sharing a Data container does not fail-stop. The retained chan_shared_node fixture prints 78end on actual native 1/4-thread runs, interpreter and emitted JS. Rebuilt 26 existing native channel/fiber consumer runs pass their success/error contracts. Repo gate passes 54/54.

Added chan_reclaim_limit.bend: 16,777,217 discarded, never-explicitly-closed handles. Original main fails with the old lifetime-limit diagnostic; after-fix native and JS print done. Single Linux samples: native 1.23s / 2,228 KiB peak RSS, emitted JS 3.10s. Neither timing nor RSS is a permanent assertion.

Tradeoff: maintainer measurements show channel-only loops 6–13% slower due to node allocation and handle dropping; there is no channel perf bench covering this cost. This is a memory-lifetime correction, not a blanket speedup. No fresh GPU, Mac, mini-cluster or ASan run is claimed for this follow-up.

[Lorenzobattistela]

Thanks, this fixes #1268: 500k buffered send/close goes from 34.5 MB to 1.6 MB here. The 17M-iteration loop no longer stops at "more than 16777216 channels", and never-closed channels are reclaimed too.

One bug: a Chan is Data, but chan_open returns a plain, non-shared ARR block. When a node holding the handle is shared and then matched, for example a +xs : List<&2, Chan(U32)> used twice, span_fade finds a non-rfc field and the program dies with runtime fail-stop. The same program works on main. Wrapping it at birth fixes it:

return rfc_wrap(e, term_blk(true, cls, loc), 1);

Please add a test for that case (a channel in a shared list cell). Also, both new tests pass on main, so neither would catch a revert of this fix.

Smaller things:

• The three one-line ifs should use braces like the rest of the file.
• The header comment still says "the rows live here".
• ASan can't see the Bend heap, where the row and nodes now live, so the "ASan clean" claim covers only the timers.
• Channel-only loops are about 6-13% slower (node alloc + handle drop per op), and no perf bench covers channels. Please mention that in the description.

[oxura]

Published a781bf7. Your d95603c arrived while I was pushing the same RFC fix; I merged and preserved it rather than overwriting the branch. I retained your tests/io/chan_shared_node.bend and removed my overlapping shared-list fixture. Your retained fixture prints 78end on rebuilt native binaries at 1/4 threads, the interpreter and emitted JS. My separate before-fix shared-list probe had reproduced runtime fail-stop, so the regression was observed, not inferred from green gates.

Added a reclamation regression that fails on original main rather than only covering old behavior: tests/io/chan_reclaim_limit.bend creates and discards 16,777,217 never-explicitly-closed handles. Original main exits 1 with “more than 16777216 channels at once” (788,152 KiB peak RSS); the patched native run prints done (2,228 KiB; 1.23s) and emitted JS prints done (3.10s). These are single Linux samples, not timing/RSS assertions in the test. The ordinary interpreter run also prints done. This checks the removed process-wide lifetime limit without a platform-dependent RSS threshold.

Rebuilt and exercised all 26 existing native channel/fiber consumer runs at 1/4 threads, including the expected closed-join and deadlock errors. Repo gate: 54/54. Added braces to the three one-line ifs and corrected the header/foreign-representation documentation.

Agreed on the remaining limits: the earlier ASan runs cover ordinary allocations such as waiter timers, not Bend heap bounds or row/node reclamation. Your channel-only 6–13% slowdown is a maintainer measurement, not a general speedup claim; I am adding both that cost and the absence of a channel perf bench to the PR description. No GPU, Mac or mini-cluster run claimed.