Seeding faults into our CI checkers

A red-team run seeded 32 small faults (flipped comparisons, off-by-one constants) into four checkers under platform/ci/: check_input_bounds.py, check_form_defaults.py, check_shared_link_clears.py and shorten_descriptions.py. Each fault was applied one at a time and the checker's own test suite run against it.

The suites caught 28. The four that escaped were each checked by hand and are equivalent:

Two flip <=/< at the edges of a regex-match range in check_shared_link_clears.py. The range starts with the literal if while the matches start with write/apply, and the range ends after else\b, where a match cannot begin. Neither edge can be hit.
Two change an index in shorten_descriptions.py: the sys.path insert position and which docstring paragraph becomes the --help text. Neither changes any output a user relies on.

No real defects turned up, so no code changed. The catch rate for these files is 100% after triage. Run records live in prompts/runs/; python3 prompts/promptctl.py measure aggregates them.