emrg: the competition template never says what decides the score has to be re-read each round #2452

What the template does not say

emrg/server/competition_prompt.md §2 ("Goal line and prize reading") tells a round to read
the prize rules and payout conditions in full, and — for a competition without prize money —
to record the metric and the current standing. It never says that what decides the final
score is itself something to go and read each round.

So a round can carry an old conclusion about how a platform turns submissions into a score,
and nothing in the template asks it to check. The four things that decide it differ per
platform and are stated in each platform's own terms:

• which submissions count — best of, last one, or a manual selection of N;
• whether the platform falls back to a default when nothing has been selected;
• the quota and when it resets;
• when the standings freeze.

The template mentions only the third, as "note the submission budget/limits if the platform
has any". A round that reads that line and stops has read a quarter of the rule.

The second half, which is worse because it decays silently

Anything set by hand once on a platform — a manual selection of submissions, a pending
queue, a freeze marker — does not follow the results that arrive afterwards. A round that
selected its N best submissions on day one keeps that snapshot for the rest of the
competition unless someone re-reads it, and every round after the first better submission is
a round scoring below what the round's own work had already achieved. The template has no
sentence about this at all, so there is nothing a round could notice it had forgotten.

Expected

The section that already owns what the score means states both halves:

1. that "what decides the score" is a per-round reading of the platform's own terms, not
   background knowledge and not a conclusion to reuse across platforms or across rounds;
2. that platform state set by hand goes stale, and that a mismatch between what is set and
   what the round's own best submissions are is that round's work rather than a note for
   later.

What the fix must satisfy

• It lives in §2, the section that owns the subject — not appended at the end of the file,
  and not stated twice.
• It names the four checks as a method, never as a table of one platform's answers: the
  template must not hard-code any platform's rule, because each states it differently and a
  constant here would be read on a platform it was never measured against.
• It is pinned by a guard asserting on the rendered prompt rather than the file bytes. A
  rule wrapped in a Jinja comment renders as nothing while the file still carries every word
  of it, so a file-bytes assertion reports a rule that no round ever reads.
• The guard says what it does not measure: whether a round obeyed the rule is a reading of
  that round's record, not something a text assertion can answer.