Two-method rule: methods sharing a blind spot are one method

[h4x0r]

PEIR-LINT-FALSE-INDEPENDENCE exists, but it is narrower than the rule it is named
for. Today it fires when one supporting node duplicates or restates another:

"{a} and {b} both support X, but one duplicates the other"

The spec's §2.1 is stronger:

A figure not reproduced by a second, differently-shaped method must be reported
with that caveat attached. Two methods that share a blind spot are one method.

The dangerous case is not restatement — it is two supports that read completely
differently and are still one line of evidence: the same tool run twice, two analyses
of the same image, two checkers vendoring the same dependency. The spec's own worked
example: a nonce test cannot see cold wallets, an explorer-label sweep can; either
alone is weak, a negative from both is strong.

peira cannot currently see this, because an Observation does not record what
produced it. Needs the Instrument node kind first.

Also from §2: an oracle is independent only if its lineage is — which is a
property of the instrument graph, not of the text of the two claims.

Source: ~/.claude/knowledge/evidence-claim-grading-spec.md §2, §2.1.