Quality cleanup: no shared-gadget mutation, faster chain build, style

[ricardojrdez]

Resolves the maintainability items raised in the post-merge code review (no behaviour change for users; output is identical).

No shared-Gadget mutation

Operation.filter_gadgets previously wrote .op/.dst/.src and side effects onto the input Gadget objects. Because the same gadget list is filtered for several operations of a chain, the last match could overwrite annotations on a shared object. It now annotates a copy via dataclasses.replace, which also gives each match fresh side-effect sets.

Faster ROP-chain construction

_build_per_comb_gadgets used to re-filter, re-sort and re-prune the full per-operation gadget lists for every register combination — O(combinations × gadgets). It now:

• sorts each operation's list once up front, and
• memoises the filter/prune result per (step, req_dst, req_src), since different combinations frequently request the same concrete registers for a step.

Output is unchanged (same filtering, sort order and pruning).

Encapsulation & style

• ArchitectureSingleton.reset() added to clear the global state cleanly (used by the test fixtures instead of poking the private attribute).
• isinstance(...) instead of type(...) == ... (operation.py, yaml_parser.py).
• Imports moved above the module constants in gadfinder.py.
• Lower bounds pinned for pyelftools / pyyaml / macholib in requirements.txt.

Tests

• New filter_gadgets no-mutation regression test.
• New test_ropchain.py (concrete, two-step, generic-register and not-found cases) to guard the chain-build refactor.
• 59 tests pass locally (3.11 and 3.13); end-to-end output verified unchanged on /bin/ls and a 32-bit busybox.

🤖 Generated with Claude Code