gate: subagent guard misses an escape that only appears after eval or a double-quoted shell body parses the text again #327

Found by the #317 differential fuzz (pre-existing; develop allows every one of these too).

The #317 fix reads a backslash before an ordinary character the way one shell parse does. Two shapes parse the text twice, so a backslash pair that is a literal backslash after the first parse becomes an escape in the second one, and the subagent guard does not see the resulting git word:

• unquoted eval: eval joins its parsed arguments and parses them again. eval g\\\it stash gives eval the word g\it, which runs git stash; eval nice\ git -C . reset gives it nice git -C . reset.
• a double-quoted shell body: inside "..." the outer shell turns \\ into \ (and \$, \", a backtick escape), and the inner shell parses the result. bash -c "g\\it stash" runs git stash.

75 of the 108 branch false-allows in the #317 fuzz (4000 cases, real sh, bash 3.2, bash 5, dash, zsh, ksh as the oracle) are this shape; the other 33 are #320.

Fix direction: for a double-quoted body, undo the double-quote escapes (\\, \$, \`, \", backslash-newline) before the body is masked; for unquoted eval, scan its argument text a second time after one round of escape removal (any deny wins). Fuzz per docs/reference/repo-gotchas.md.