research: add SYM-ARCH-002A4 adversarial construct-validity controls

[Tristan-Stoltz-ERC]

Summary

Adds SYM-ARCH-002A4 adversarial construct-validity controls from issue #55.

This is a draft stack on #59 (research/sym-arch-002a3-validity-v1). It consumes A3's generated-dataset/oracle validity layer and asks a stricter question before any architecture score counts as evidence:

Can a scientifically trivial train-only shortcut solve the benchmark anyway?

If yes, the result is INCONCLUSIVE_BENCHMARK, not an architecture win/loss.

No architecture capability result is produced by this PR.

Train-only shortcut controls

A4 reports and gates:

• training-majority predictor;
• single-feature marginal predictor;
• exact feature-assignment lookup;
• deterministic categorical nearest neighbor;
• prevalence-preserving shuffled-label single-feature controls;
• prevalence-preserving shuffled-label nearest-neighbor controls;
• executable symbolic-oracle positive control;
• deterministic fair-chance sanity control.

The single-feature model chooses its feature using training accuracy only. Evaluation labels are used only for final scoring.

Fitted shortcut models consume learner-visible features only. example_id, support_tags, support metadata, and oracle outputs are excluded from fitted model inputs.

Finite-sample resolution gate

A4 computes the 95% Wilson upper accuracy bound for a fair Bernoulli classifier at the actual evaluation-set size.

If ordinary chance uncertainty can reach the tighter frozen shortcut ceiling, the benchmark is automatically INCONCLUSIVE_BENCHMARK and must increase evaluation support.

This intentionally catches coarse held-outs such as four-item evaluation splits before anyone treats a one-item difference as a precise architectural effect.

Fail-closed contract

A4 returns only:

• passed;
• inconclusive_benchmark.

It becomes inconclusive when:

1. A3 structural/oracle validity fails;
2. evaluation support cannot resolve the frozen ceiling;
3. the executable oracle misses its frozen floor;
4. majority, single-feature, exact lookup, or nearest neighbor reaches/exceeds the frozen shortcut ceiling;
5. mean shuffled-label single-feature or nearest-neighbor accuracy reaches/exceeds the frozen shuffled-control ceiling.

A benchmark failure is never converted into a negative architecture claim.

Shuffled-relation controls

Training labels are deterministically permuted while preserving prevalence. At least four unique frozen shuffle seeds are required. A4 records both mean and maximum accuracy across shuffles; v1 gates on the mean to avoid making validity hinge on one lucky permutation.

Tests

Focused tests cover:

• a clean XOR-style compositional benchmark passing low-complexity controls;
• an injected single-feature label channel becoming inconclusive;
• a four-item evaluation split failing the finite-resolution gate;
• label-correlated support metadata leaving fitted shortcut scores unchanged;
• prevalence preservation under deterministic label shuffling;
• rejection of weak/duplicate shuffle contracts.

Threshold hygiene

The 0.80 ceilings used by unit tests are synthetic fixtures only. They are not scientific defaults.

Claim-bearing use must freeze before result inspection:

• shortcut ceiling;
• shuffled-label mean ceiling;
• oracle floor;
• chance seed;
• complete shuffle-seed set;
• A3 policy;
• generator/version and seed manifest.

CI

Adds SYM-ARCH-002A4 Construct Validity with:

• targeted rustfmt check;
• focused construct-validity tests;
• psych-bench library check.

The unrelated stale LoopTrialResult::cycle_reward fixture repair has already been propagated identically across #53/#57/#58/#59/A4 so this stack does not reintroduce that known compile failure. It does not appear in this PR's scientific diff.

Wording ceiling

A pass supports only:

The benchmark passed the implemented v1 structural/oracle and low-complexity shortcut controls under the frozen policy.

It does not establish that the benchmark is universally shortcut-free, and it does not establish Symthaea superiority.

Depends on #59. Tracks #55.