Stricter safety checks for declare, typeset, local, export, readonly, and unset commands

What

When Claude Code decides whether a shell command is simple enough to run without invoking the full shell, it now rejects more edge cases that could behave differently than expected:

• declare, typeset, and local commands are rejected as too complex unless their operand is a plain variable name with a simple, inert value, since the real shell could otherwise treat it as an array subscript, an expanded name, or a compound assignment.
• export and readonly commands are now also rejected when their operand looks like an array assignment (matching flag patterns the shell would evaluate as a compound array assignment).
• unset no longer just checks for a bare variable name; it also allows a second, more permissive pattern under certain conditions, but anything that still doesn't match either pattern is now rejected as too complex, since the shell might expand or evaluate it unexpectedly.
• These rejections are now marked internally with a differential flag, replacing a fixed classification that was previously used for these cases.

Why

These checks exist to stop Claude Code from treating a shell command as "simple" when the real shell could actually parse or evaluate it in a surprising way (for example, expanding into an array assignment). Tightening them reduces the chance that a command behaves differently than Claude Code's analysis assumed.