[sjkim1127]

Summary

• Supply the bounded x86-64 Linux HLE runtime guard/header contract from guest AT_RANDOM before main; keep existing TLS and static/interpreter/shared-library initialization intact.
• Handle __stack_chk_fail as nonreturning termination rather than a fake successful return.
• Make sandbox --json a complete JSON stdout document, with captured stdout/stderr bytes in the typed metrics report; reuse existing echo control and VFS storage.
• Add owned JIT fixtures/regressions, a generic candidate-validation workflow and an evidence-based public crackme walkthrough.

Fixes #198
Fixes #199

Measured acceptance

Public Reversing-ELF crackme4 was solved from Fission string context, callgraph and disasm without source/solution text. Original + symbol-stripped copy, each with one correct and four incorrect/boundary inputs: 10/10 exact stdout, stderr and exit-status matches against Linux/amd64 process oracle. No input mutation or opcode/HLE/unknown-syscall/unhandled-userop misses. Both branches exit 0, so output is explicitly treated as the verdict.

Baseline halted on the unmapped FS guard read; its guest message also polluted JSON stdout. Candidate runs now halt cleanly at 417/419 instructions with structured output.

Validation

• Focused nextest: 10 passed.
• Full emulator + CLI nextest: 379 passed; 3 pre-existing ignored corpus/throughput tests.
• Crate check/clippy with CI policy, locked release build, workspace fmt/diff checks and Python compilation passed.
• Existing CLI/project/history smoke: 43 commands + 2 unchanged binary hashes, fresh DB directory.
• JIT corruption test, alternate ABI rejection, existing-pointer/non-HLE preservation, modified AT_RANDOM seed, exact non-UTF-8 capture, Linux libc/syscall + Windows console JSON/text and metrics-file equality covered.

Limits

One beginner challenge plus its derived stripped copy, not independent samples or general crackme coverage. No automatic solving, ELF32 parity, packing/anti-debug support, full PT_TLS/DTV/pthread state or guest SIGABRT delivery. The failure HLE uses conventional status 134. Console capture is retained VFS descriptor content, not a redirection transcript. Docker is a validation oracle only; production remains Rust-owned and vendor-independent. No decompiler algorithm or ranking claim. Prior decompiler work and PR #176 preserved.

Report: docs/validation/2026-10-12-crackme-workflow.md
Walkthrough: docs/CRACKME_WORKFLOW.md