Do the 55 'feasible' alloys survive once two discredited checks stop being trusted?

E131 — Do E117's 55 feasible compositions rest on two withdrawn instruments?

The audit of the record raises this as the highest-blast-radius open problem, and it is not in the task queue. E117 defines feasibility as two conditions:

p_survives_window > 0.9      AND      orders_in_window false

Both instruments have been withdrawn, and neither withdrawal reached screen().

• E99 withdrew the ordering gate as a clearance: "A nonzero T_order is a real signal ... A zero is not a clearance." It measured the failure curve off equiatomic - x_major 0.56 -> ratio 0.72, 0.69 -> 0.21, 0.73 -> 0.00 - and the error is one-signed: the gate always under-reports. E118's Pareto front is almost entirely 0.75/0.25 binaries, squarely in the regime where the gate was measured returning zero against a real transition.
• E92 falsified the Q proxy, over-predicting by up to 1.58 eV and "always in the direction that manufactures stability", with the error growing monotonically with element count. E93 then stated the remedy: "The term stays out of the screen."

If both hold, both halves of "feasible" are broken in the same direction, and E118, E122, E123 and E125 all build on the 55.

Predicted:

1. The Q term is still in screen() and still feeds p_survives_window - i.e. E93's stated remedy was never applied. Verifiable by reading the code.
2. A large share of the 55 carry T_order exactly 0, which E99 says is not a clearance. I expect more than half. If it is under a tenth, the withdrawal barely touches this set and the problem is smaller than the audit claims.
3. Refusing zeros as clearances shrinks the feasible set substantially - to fewer than 30 of the original 55.
4. The compositions lost are disproportionately the extreme ones (0.75/0.25 and beyond), because that is where E99 measured the gate failing, while near-equiatomic ones survive.

Falsified if the zeros are rare or evenly spread across composition extremity, in which case E99's withdrawal, though real, does not undermine E117 and the 55 stand on the ordering side. The Q side would still need separate treatment.

E131 result: all four predictions confirmed. E117's feasible set collapses from 55 to 12, and only one of those twelve is cleared by an instrument operating inside its validated regime.

Prediction 1 confirmed. Q_PER_KB_TMELT = 18.4 at verifier.py:90 feeds q_est (212), p_happens (229-231) and p_survives_window (294). E93 wrote "The term stays out of the screen." It is in the screen. The remedy was stated and never applied.

Predictions 2, 3 and 4 confirmed:

of the 55 "feasible" compositions
  T_order exactly 0 - which E99 says is NOT a clearance        43   (78%)
  T_order nonzero and outside the window                       12

the axis E99 measured the gate failing along
  zeros:    mean x_major 0.571,  23 of 43 at x_major >= 0.70
  nonzero:  mean x_major 0.368

The zeros sit exactly where E99 measured the gate returning 0.00 against a real 649 K transition. Cr0.75 Ni0.25 - the top of E118's Pareto front on every axis the ladder can compute - is one of them, at x_major 0.75.

The twelve survivors are not twelve results. Eleven of them have T_order between 18 and 81 K - that is, just below the window's 90 K floor - reported by a gate whose error E99 measured as one-signed and always downward. A reading of 25 K from an instrument that returns 0.00 for a real 649 K transition is not evidence of a transition below the window. Exactly one composition, Mo0.50 Ta0.50 at 1655 K, is equiatomic - where E99 says the gate is reliable - and far enough above the window that the one-signed error cannot reach it.

55 nominally feasible  ->  12 surviving the withdrawal  ->  1 robust

What this withdraws:

• E117's "55 feasible compositions" and its "3.2 per cent of the grid" figure. Refusing zeros gives 0.7 per cent.
• E118's Pareto front, whose top three entries (Cr0.75Ni0.25, Co0.33Cr0.33Ni0.33, Co0.25Cr0.25Cu0.25Ni0.25) include two zeros; the oxidation comparison that produced "28 times better" is drawn from a set that does not survive.
• E122's "10 of 55 flagged" - the denominator is wrong.
• E125's "E117 already established that only 3.2 per cent of the grid passes", which was used to overturn the archive-arm framing. The corrected figure is 0.7 per cent, which strengthens rather than weakens that particular argument.

What survives, and it is the load-bearing half of E117. The enrichment claim is unaffected and slightly stronger: refusing zeros, the whole grid passes at 0.7 per cent and the best 50 by reward at 10 per cent - a 14-fold enrichment against the 12-fold reported. So E117's actual conclusion - that the search is not broken and the reward is a good filter - stands. What collapses is the absolute size of the feasible set and everything built on the particular compositions in it.

And the Q side is untouched by this entry. p_survives_window remains the other half of every feasibility verdict above, and it is computed from an estimator E92 falsified as over-predicting by up to 1.58 eV "always in the direction that manufactures stability". The 12 have not been re-tested against that; they are what survives the ordering withdrawal only.

231 passing.

p_survives_window > 0.9      AND      orders_in_window false

of the 55 "feasible" compositions
  T_order exactly 0 - which E99 says is NOT a clearance        43   (78%)
  T_order nonzero and outside the window                       12

the axis E99 measured the gate failing along
  zeros:    mean x_major 0.571,  23 of 43 at x_major >= 0.70
  nonzero:  mean x_major 0.368

55 nominally feasible  ->  12 surviving the withdrawal  ->  1 robust