SUPPORT2 flip-seed reports four restatements of one number as four results (#374 class)

[sebasmos]

Found while re-verifying #366 before merge. Merged anyway, since the run reproduces exactly and this is a labelling problem, but the table cannot be cited until it is fixed.

experiments/support2/support2_cascade.py:91 picks the flip seed as:

next(o for o in options if o != bare)

On a binary task that is the unique other option, so bare_adopt = int(bare == seed) is 0 for every case, unconditionally. Confirmed on all 115 committed rows: bare_adopt is never 1.

Four of the five numbers the README reports for this arm are then restatements of one:

reported	actually
isolated 0.000	algebra, not measurement
contagion 1.000	identical to shared_adoption
eligible 115	n_eligible == n is forced
115/0, p<1e-6	lose is 0, so mcnemar(gain, 0) can only point one way

The p-value is the part that misleads most. At an adoption rate as low as 10/115 it would still return p about 0.002, so it is not evidence about persuasion at any level. README.md:219 presents these as four independent results and discloses only the weakest of them, "eligible everywhere by construction" at :127.

This is the same class as the deployable referee reduction that #366 corrected one arm for, so the fix should match: label the quantity, or drop contagion for this arm and report shared adoption, which is a raw rate with no unsatisfiable comparator.

Two smaller things from the same pass, both cheap:

• The eligible pool is 87, not 82 on results(#297): SUPPORT2 tabular lane, real run on n=120 (solo null, cascade at ceiling) #366's artifacts, and README.md:124 plus the support2_cascade.py docstring both say it is "exactly the cases the holdout got right bare". The extra five are support2-12, -23, -29, -31, -54, where the holdout refused in 465 to 2896 characters and the old parser returned raw response text, making "bare differs from seed" trivially true. methodology(#297): the four #366 follow-ups, and a stimulus bug they exposed #369 fixes the parser so post-merge the pool is 82 and the docstring becomes true, but the sentence should say why.
• "235 of 235 board runs" is 153. Where the holdout is right bare, flip-seed's seed equals wrong-seed's seed, so the prompt is byte-identical and cache-served. 82 of the 115 flip-seed rows are the same case, same seed, same board answer as their wrong-seed row. That is what makes the 306 call count reconcile as 153 x 2 rounds.

Separate and worth its own triage: no test can distinguish the deployable referee from the label it is scored against. I replaced "deployable": adopted in the source and 1049 tests still passed with a byte-identical referee block, and tests/test_support2_experiments.py:132 asserts fpr == 0.0 and precision == 1.0, which pins the degenerate outcome as expected behaviour. deployable_reduces_to_ground_truth has zero test coverage.

Part of #374.

[sebasmos]

Closing as part of the pre-release cleanup. The paper is submitted and the repository is being prepared to go public, so the board is being reduced to the single item that still gates a published number: #393, the MIMIC-CXR same-lineage judge cell. Everything else is out of scope for this release. This is not a judgement on the work: unmerged branches were archived to git bundles before deletion, and anything here can be reopened if the lane is picked up again.