emrg: competition_prompt.md says to read a probe, not what to read it against #2504

What is missing, and where

emrg/server/competition_prompt.md §Phase D gives four steps for establishing the right answer for a class of cases before changing what produces them. Step 3 reads:

Read the probe's outcome as that class's answer. A probe whose result is not read has bought nothing; a class whose answer is established is a fact the remaining cases can be judged against.

It says the outcome must be read. It does not say what the outcome is compared against.

Why that leaves the reading undecidable

When a probe's score is compared with the best submission so far, a zero delta carries two opposite meanings at once: the cases the probe was meant to restore came back, or the probe changed nothing that mattered. Both print the same number, so the submission bought no fact, and the round cannot tell a class whose answer is established from one it has not touched. The four steps read as complete without this, which is what makes it easy to leave out.

An absolute floor removes the ambiguity. Submitting a deliberately wrong answer — in the same format as a real one, so the platform grades it rather than rejecting it — pins a scoring floor, and every later probe is read as score - floor. On a per-case grader that difference answers zero or the case's own value, which is a yes/no question per probe rather than a number with two causes inside it.

What the fix must satisfy

• §Phase D names the baseline a probe is read against, in the paragraph that already owns the subject, without restating the four steps.
• It says the floor submission must be format-valid, and gives the reason: a rejected submission provides no floor.
• It states why a zero delta against a best-so-far baseline is undecidable, so the rule is reasoned rather than asserted.
• A test pins both phrases, in the style of test_phase_d_establishes_the_answer_on_a_sample_before_changing_the_class.