[helly25]

Preserve the cross-platform evaluator experiment and its rejection so future optimization work can build on the measurements. The worker-wrapper allocation saving does not justify the retained Linux regressions; shipping runtime stays unchanged.

AG;DR

• Inspect the actual ARM64 ThinLTO matcher call site: a worker-vtable evaluation call remains. Do not count test-only qualification executors as production overhead.
• Retain a complete two-file candidate replacing allocated production worker wrappers with direct observer/private-worker storage. Keep the heap-stable prepared plan and existing test-only qualification-state alternative.
• Both hosts pass four focused release suites covering whole-run differential behavior, moved owners, capture isolation, pool transitions and output. Isolated ASan diagnostics execute 108 selected production/tree phase cases per variant per host. Coordinator worker creation drops from one allocation to zero; concurrent creation saves one allocation and a 16-byte request. Steady-state allocation counts for those examples were already zero.
• Measure 34 native cases x three sessions and 30 in-memory cells x three sessions on macOS ARM64 and Linux x86-64 drogon. Retain nine rounds/fastest-seven means, rotating participant order, Linux CPU affinity, native output oracles, complete source/binary identities and raw samples. Compilation and diagnostics do not overlap timing on a host.
• Confirm six native cases over five sessions and seven kernel cells over three sessions, adding a participant copied byte-for-byte from baseline. Linux early-difference/four-worker comparison is +40.18% initially and +14.38% in confirmation. Native confirmation also retains +4.40% for the one-worker large comparison and +2.81% for 1,000-file native line output. macOS mostly remains near baseline. Preserve noise and reversals; the exact cause is not established.
• Reject this implementation and keep it in the lossless evidence archive. Update the existing O12 checklist, performance analysis and release-site report mapping. A deeper typed-plan design remains unqualified; O13-O22 and existing EP/JC work continue independently.

This PR changes documentation and retained evidence only. No runtime, parser, flags, safety policy or worker scheduling is changed. Normal enforcing hooks and current pushed-head CI apply to the documentation PR.

Based on merged #1051 (d903ea98a98f4220d9372ab0404ee974eccb8690). Synchronization retains exactly the previously validated documentation tree; the current main-context CI run is authoritative.