[NeverSightAI]

Cached no-return lookup could disagree with live image ownership, synthetic object tail callees were rejected as non-executable, and exception parents without an ordinary RETURN could lose their catch result by being classified as void. This change preserves those contracts at their shared owners and fixes the cross-platform failures exposed during performance qualification.

• Match cached/live veneer admission, exact IAT precedence, section aliases and tagged Thumb targets. Retain loader-owned callable object externs across ELF, COFF and Mach-O while refusing data slots as code.
• Require the existing no-return proof before an absent ordinary return can imply void. External catch continuations retain their results through both C routes.
• Execute all five Win64 variadic arguments through HighC and LLVM-C, including the matching Win64 test-callee ABI on non-Windows AArch64. Preserve declared argument counts and partial-write refusal.
• Correct stale unallocated-instruction, resumable-trap and overlapping Mach-O entry fixtures; retain bounded generated-C compiler diagnostics on slower Windows runners.
• Run complete LibC and return-contract owners plus the real worker call-site test in native CI with TBB enabled and disabled.
• Complete the 57 missing source messages in all 11 GUI catalogs. Fix missing command icons/menu entries, start searches after pending navigation settles, preserve terminal MCP errors when stopped processes deliver late stdout, keep mock SDK filesystem paths Unicode-safe, and destroy views that borrow window state before that state is released.

Linux Release validation: full Lift 7,489 passed, zero failures; 34 Darwin-native cases require macOS. LibC 71/71, session 193/193 with matching optional inputs, pipeline 49/49, return contracts 21/21, HighControlFlow 430/430, store forwarding 28/28, worker/allocator 47/47. Swift and real AArch64 parity coverage were enabled. Format, documentation and 51 CI policy tests pass. Native Linux and Windows clang-cl artifacts confirm the critical owners without skips in both scheduler configurations. For MSVC-built SDKs, generated-C execution fixtures select Clang, matching the native unit fixtures; this validates the SDK and emitted semantics without claiming that every Clang C extension compiles directly with cl.exe.

Qt 6.8.3 Release: all 23 GUI/worker CTests pass locally on the final lifetime fix, including the unchanged search and Unicode drag/drop assertions. The late-output MCP regression first failed locally, then passed 20 complete repetitions. The prior shutdown failure was reproduced as a use-after-free under ASan. The exact bookmark and translation cases pass after the lifetime fix; the offscreen focus regression also passes in eight separate ASan processes. The complete ASan controller run passes all 192 cases with no sanitizer reports. Windows offscreen CI explicitly loads the system font directory: a font-free local run reproduces its exact 407-versus-612 width failure, and real fonts satisfy the unchanged assertion. All 11 catalogs compile with 1,154 matching source keys, placeholders and unique menu access keys.

The comparison engine is now the supplied IDA 9.5 Beta1 (kernel 9.5, build 261001). Seven frozen input images use identical explicit address lists: 3,198 requests, 2,977 common successes, with all refusals retained (8 NeverD; 213 IDA). NeverD leads in six of seven common-success warm decompile aggregates. Three fresh sequential pairs confirm Android32 remains 16.8% slower by median, while MSVC x64 is 3.7% faster by median. The library open, analysis, warmup and full-request total is lower for NeverD on all seven images. Startup, memory, individual times and errors are recorded separately; the total-path measure is not GUI startup. This replaces the previous IDA 9.4 comparison claim. A 2,298-request source comparison retains the success/refusal sets; only three return signatures change to conservatively retain unknown values.

Current-head native macOS, Windows and remaining CI qualification are in progress.