fix: plain parallel blocks take the replace-tree's duplicate semantics

[ronaldtse]

Two selection regimes, both measured against Ruby:

• An all-plain parallel block compiles to a replace tree, where a later duplicate from overwrites an earlier one — masm-mon lists sub "i" "й" … sub "i" "и" precisely so the last wins (measured with a synthetic duplicate map: i → и, ab → Y).
• A block containing guards, or any rule whose from or result holds a boundary-like token (Ruby raises "Can't use boundary in a par context", found at child 65 of bgnpcgn-bal's parallel), falls back to the megaregexp, where the earliest equal-key rule wins — bal's و → o, not the later w.

The engine now drops every duplicate plain pattern except the last when the block is par-safe (expr_is_par_safe), and keeps index order otherwise.

Direct corpus sweep: 46 failures / 14 maps → 35 / 13 — masm-mon (11 examples) fully healed, bal restored. Suite 46 passed / 1 xpassed.