D003 boundary correction
------------------------
If a plaintext action permutes deck positions and then emits the top card,
two identical action sequences preserve a fixed correspondence between the
cards in their two starting states. That correspondence includes the top card
BEFORE the first shared action as well as every output after it.

A ciphertext isomorph of length30 therefore naturally motivates29 shared
actions connecting its30 observations. Equating the first producing action
as well is an additional boundary assumption. In this corpus that stronger
assumption conflicts with a preceding-output correspondence, as the parent
and independent reviewers are documenting separately.

The corrected endpoint hypotheses are:
  E4[69:98] = W4[72:101]       (29 actions)
  E4[69:98] = E5[70:99]        (29 actions)
  W1[35:52] = W1[65:82]        (17 actions)
Indices are zero-based and slice ends excluded. These remain hypotheses about
plaintext actions; the equalities are not recovered text.

Rerunning the endpoint screen changes neither necessary pair count:
  Primary E4/W4 hypothesis:322 survivors of729 baseline pairs.
  All three corrected contexts:35 survivors.

The compact reason is still present: W1[38] and W1[68] lie inside the corrected
third context and must use the same action. Their consecutive return gaps are
8 and9. That ONE equality alone leaves the same35 endpoint-compatible pairs
among all3321 possible off-top pairs, without using an alphabet-size bound.
See T002-C005 and experiments/T002-E004/verify_gap89.py.

Earlier stronger-assumption results are preserved in T002-E003 and C004 with
an explicit correction caveat. They must not be silently presented as results
under these weaker, boundary-trimmed hypotheses.
