P2: UNSAFE backend scripting escape hatch (eval/exec)

[mahaloz]

What & why

The universal escape hatch for backend-specific work the abstracted API doesn't cover (IDALib's py_eval/py_exec_file). Kept explicitly unsafe and backend-specific — not portable, not sandboxed.

New capabilities

• decompiler eval "<expr>" and decompiler exec "<code>" | --file <path>.
• DecompilerInterface.backend_eval runs Python inside the backend process with the interface bound to deci: idaapi is importable, deci.flat_api (Ghidra), deci.project (angr), deci.bv (Binary Ninja). One generic base implementation serves every backend, routed through the server's main-thread dispatcher so it's safe on IDA/idalib.
• eval returns the expression's repr; exec captures stdout and a result variable. Errors return a traceback and exit non-zero.

Example

$ decompiler eval "deci.name"
'ida'
$ decompiler eval "len(list(deci.functions.keys()))"
13
$ decompiler exec "print(hex(deci.binary_base_addr))"
0x400000
$ decompiler exec --file ./my_ida_script.py
$ decompiler eval "1/0"        # exits non-zero, prints traceback

Design note

The failure field is named traceback, not error, on purpose: the client/server wire protocol reserves a top-level error key for RPC failures, so a result dict with an error key made every successful call look like an RPC error. Renaming the key fixed it.

Tests

test_eval_expression, test_eval_has_backend_access (verifies deci is bound and the backend is reachable), test_exec_captures_stdout, test_exec_result_variable, test_eval_error_is_reported. All pass on IDA, Ghidra, and angr — the mechanism is backend-agnostic.

---

Eighth and final PR in the IDALib-parity series. Based on feat/cli-patching (#210).

This completes the gap-closing set: persistence, comments, globals/signatures, typed reads/addresses, search/imports, define/undefine, patching, and this scripting hatch.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HfZHdprXg38re1XyKxzGg5

[mahaloz]

Integrated into main as commit b8ad5e4. The stack got tangled when the bottom PRs (#204/#205) were squash-merged (#205 landed on feat/cli-persistence instead of main), so I reconciled by cherry-picking the whole series onto main linearly. The exact reviewed content of this PR is on main — closing this stacked PR as its changes are shipped.