research-result: forbid positive confirmatory claims from being supported only by Missing/NotComputed outcomes

Context

Follow-up to PR #159 (symthaea-research-result) and generic research-integrity prerequisite for #1905 / #1898.

The current result layer correctly requires every preregistered primary metric to appear as an explicit outcome, including Missing { reason } and NotComputed { reason }. That is good.

However, the current claim validation checks only that a referenced metric id exists in the result manifest. It does not require the referenced metric to contain an observation capable of supporting the claim disposition.

Therefore a structurally valid manifest can currently express the equivalent of:

metric M = Missing("sensor failed")
claim C = ConsistentWithHypothesis
claim C references M
claim C interpretation = Confirmatory

The metric did not disappear, but the positive claim is still unsupported by the recorded observation state.

Core invariant

metric present
!=
metric observed
!=
metric supports positive/negative claim

A Missing or NotComputed primary result must never, by itself, support a confirmatory ConsistentWithHypothesis, InconsistentWithHypothesis, or NullResult claim.

Proposed evidence-state semantics

Classify MetricOutcome into an explicit support state, conceptually:

ObservedNumeric
ObservedBoolean
ObservedCategorical
Missing
NotComputed

and define conservative claim compatibility.

At minimum:

• ConsistentWithHypothesis requires at least one referenced observed metric or an explicitly typed observed/result artifact allowed by the frozen analysis plan;
• InconsistentWithHypothesis likewise requires observed evidence;
• NullResult requires observed evidence sufficient to apply the preregistered null/no-effect decision rule; absence of measurement is not a null result;
• Inconclusive / NotEvaluated may be supported by Missing / NotComputed as appropriate;
• DescriptiveOnly should not be promoted into confirmatory hypothesis support merely by citing an analysis artifact.

Do not hard-code domain statistics into this generic crate. The generic layer should only enforce the distinction between observed evidence and absent/uncomputed evidence; domain/analysis code decides whether an observed value meets a preregistered criterion.

Primary metric completeness vs evidential adequacy

Preserve the current good rule:

all primary metrics must appear

but add a separate rule:

claim evidence must be semantically adequate for its disposition

These should remain distinct.

A campaign may therefore be procedurally confirmatory while yielding an inconclusive claim because a primary observation is missing. Confirmatory describes protocol lineage/analysis status; it must not imply that the scientific hypothesis received positive support.

Claim evidence projection

For each ResultClaim, derive a small evidence summary such as:

observed_metric_ids
missing_metric_ids
not_computed_metric_ids
observed_artifact_ids

Then validate disposition compatibility.

If artifact-only confirmatory hypothesis claims remain supported, require the artifact kind/role to be explicit and bound by the frozen analysis plan rather than treating any arbitrary digested artifact as evidentially interchangeable with an observed metric.

Adversarial regressions

At minimum:

1. ConsistentWithHypothesis citing only Missing metric -> reject;
2. ConsistentWithHypothesis citing only NotComputed metric -> reject;
3. InconsistentWithHypothesis citing only missing/uncomputed metric -> reject;
4. NullResult citing only missing/uncomputed metric -> reject;
5. Inconclusive citing an explicit missing primary metric -> allow;
6. NotEvaluated citing explicit not-computed metric -> allow;
7. observed numeric metric may support a claim subject to existing protocol/reference checks;
8. one observed + one missing metric remains explicit; missing evidence is not silently ignored in reporting;
9. arbitrary Analysis artifact alone cannot launder a missing primary metric into a positive confirmatory claim unless an explicit frozen artifact-evidence rule allows it;
10. imported/deserialized claim/outcome combinations receive the same semantic validation as constructor-created records.

Relationship to #1946

#1946 closes chronology/completeness/current-head integrity for amendments/deviations and recursively validates imported research evidence.

This issue is complementary: even a complete, correctly ordered result lineage can still make an evidentially invalid scientific claim if Missing is treated as though it were an observation.

Acceptance gate

• metric presence and metric observation are separate concepts;
• positive/negative/null hypothesis dispositions cannot be supported only by missing/uncomputed outcomes;
• missing/not-computed primary metrics remain first-class reported results;
• procedural confirmatory status does not imply positive hypothesis support;
• artifact-only hypothesis claims have an explicit conservative rule;
• claim compatibility is recursively validated on imported manifests;
• exact-head tests land in #1940 qualification scope.

Non-claims

This does not decide whether an observed effect is scientifically meaningful, statistically significant, causal, or generalizable. It prevents absence of observation from being represented as positive/negative/null evidence.