[cpunion]

A native program can abort while importing retained poll readiness if another producer has already published readiness or closing for the same operation. The unchanged old net/http/httptest image reproduces SIGABRT; failure-only debugger captures show PollOperationPostDuplicate(2) rejected by the native wait adapter before the strict entry validator reports Invalid.

Accept the existing coalesced result and return to the reducer, preserving the first mailbox result. Invalid, stale and retired operations remain rejected. The native entry validator, protocol and G/P/ABI layouts are unchanged.

Validation:

• A deterministic regression uses real POSIX poll revents and exact operation IDs, forces Ready-first and Closing-first publication, and checks one delivery, preserved Closing, and recycling. Restoring only the old comparison fails both orderings; the fix passes ten shuffled race-enabled runs and rejection controls.
• The rebuilt frozen fixed httptest image passes 2000/2000 consecutive P=2 executions with an unchanged hash. The original old-image failure is retained.
• Frozen production 828bfe0 passes compiler/SSA/protocol/FFI, 73 source/fleet checks (26 linked native fleet cases), runtime core/doorbell race tests, 2462 FD/Open executions, 315 reflected positives and 18 intentional negatives, and 9 independent archive tests (20 images, 60 executions). Full test/go passes 276/276 with no failures or skips; all 45 nested compiler process domains joined.

Host follow-up 1613848 updates two source-layout assertions after extraction of the import helper. CI follow-up 1020ee0 runs all four archive consumers in fresh processes under the unchanged 6 GiB virtual cap, 2 GiB Go soft target and 12m/11m limits. All four pass; the original cumulative BoundLarge OOM and failed images remain retained.

CI 39105987770 at 1020ee0 finishes 19 jobs passing and one genuine timer-fixture failure. Failure-only captures and a collector mark watch identify libgc's static mmap endpoint hint GC_unix_mmap_get_mem::last_addr conservatively marking a public Timer at an adjacent page start, which then retains its channel. The exact untraced image also reproduces locally on run 4203. Five matched no-write debugger controls fail at the original five-second assertion; clearing only that allocator hint lets five matched runs pass in about 90ms. Diagnostic runs: failure capture, mark source, causal controls.

Test-only follow-up 6fcca26 selects the same 32 Timer/Ticker pairs away from that native page-boundary false root and stops rejected guard pairs. All 128 selected weak-reference checks, the five-second deadline, and full manager retirement assertions remain identical. Added retained-Timer and extra-G negative controls independently exercise both failure assertions. Native Linux validation passes P=1x100 and P=2x1000; both controls produce their exact expected panic at five seconds. This corrects the fixture's collector assumption; conservative GC false roots and lack of a general prompt-weak-collection guarantee remain explicit boundaries. It does not change production GC or timer ownership.

PR merged into cpunion/llgo:llvm-coro as 603c781 after exact-head CI 39111825084 passes 20/20, local native/WASI suites pass all three timer fixtures each (including both native negative controls), and fresh reviews/inline comments/issue comments/review threads are empty. The merged file tree matches the tested head.

The new 1082-case GOROOT campaign, 73 executable demos, 180 stdlib packages and refreshed performance/size measurements remain separate overall-acceptance work. Historical R6 results are not counted as runs of this production fix.