Fold what the gate conversion taught into the hardening rules

[haksungjang]

Eight things the ER66 gate conversion turned up, folded into rules 7 and 8. No
new rules, and nothing that could not be attached to something already there.

Rule 7, in the "a tool found nothing" bullet, which already says a pattern
finds the notation you imagined while a parse finds what is there:

• Reaching for a parse is half of it. The census searched ast.dump() output for
  "alembic" in double quotes and reported zero, because dump writes
  apostrophes. The string matching had moved up a level, not gone away.
• What a comparison normalises away is its definition of "the same", and that
  decision produces the answer, not the comparison after it. Erasing string
  literals merged eight unrelated fixtures, because for a fixture the strings are
  the payload; erasing decorators would have merged fixtures with different
  lifetimes into one recommendation.
• Widening what a census counts invalidates the sample taken before it. Widened
  to a second way of reaching the database, it swept in 83 modules that skip over
  a missing binary and never open a connection. It was wrong six times in all,
  five of them immediately after the definition grew.

Also a new bullet for a way an assertion can hold nothing: a rule induced from a
single example keeps that example's accidents. An exemption list with one entry
asserted that exempt modules have no gate, which was true of that one by
coincidence; the next three keep their gates on purpose and the assertion failed
for all three.

And the counterpart to "a specific wrong answer is more dangerous than an empty
one": make a tool say what it does not do. tools/static-checks names the CI
steps it deliberately omits and why, and its contract test is what noticed a
later change adding a CI step without registering it - not review.

Rule 8, which already says to check a surviving mutation actually landed:
the same question applies before a bulk change. Declare how many files an edit
must touch and stop with nothing written on a mismatch, since a pattern matching
nothing reports "nothing to do" and that reads as "not applicable". Over 193
modules it stopped three runs, all three the operator's arithmetic rather than
the tool - including wc -l twice on a list with no trailing newline - and never
once the transformation. Worth saying which input was weaker.

A declared number can also disagree with you: a prediction that a batch would
shrink by one came back unchanged, and without it the correct result would have
confirmed a wrong model. Keep the count independent of the tool, or the two
always agree and nothing is checked. Hold the same bar whatever the size - one
file edited by hand is not safer than a hundred edited by a transformation
already run against the whole tree. And two identical failures mean the thing
being changed is not the variable: three attempts to repair a docstring each
moved the insertion point and each failed the same way, because the fragment
being inserted carried a stray terminator.

EN and KO move together.