test(runner): prove the operator-bind mode-to-mask decision both ways

Codecov flagged two uncovered lines in `applyPolicy`, and the gap it found is
real: every call site in this tree passes an EMPTY assignment, so the loop that
reads an operator bind's mode had never run in any lane. Both arms of that
mapping were unproven — and they were wrong once already, when bwrap mounted an
assigned bind and landlock denied it, leaving every lease on that runner
reading an assigned path as absent.

`accessForBindMode` makes the mapping a pure decider beside the sets it reads
from, unit-tested on the property rather than the constant: a `read_only`
assignment can never carry WRITE_FILE, a `read_write` one always does, and
`read_only` keeps EXECUTE because operators bind tool and model trees at that
mode and a mask without it makes them useless.

A Linux fork proof then shows the decision survives the syscall — an assigned
directory opens for reading at `read_only`, a file is CREATED inside one
assigned at `read_write` (creating, not opening: the weaker half would pass
either way), and a sibling assigned to nothing is refused. That third path is
what makes it a proof: without it a ruleset that restricted nothing at all
would satisfy every other assertion. Built under `/var/tmp`, which no baseline
list, floor or tmpfs names — under `/tmp` the writable floor would grant it
regardless of any bind.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>