feat: publish Claude oracle-disagreement flagship - #124

What changed

• publish a provenance-bound oracle-disagreement receipt for validate-agent.sh: don't abort at the first warning (set -e + ((x++))) and stop false-flagging valid agents anthropics/claude-code#89404
• bind the existing claim matrix to ORACLE_CONTRADICTED_BY_PRODUCT
• promote the Reality Lab entry from an open probe to a published contradiction artifact
• add a dedicated Reality artifact documenting the submitted-judge / product-oracle split
• add a second README flagship boundary beside Bluefin
• add regression coverage for receipt binding, candidate identity, and explicit external-oracle provenance

Why

Bluefin is CounterProof's positive flagship: same oracle, controlled intervention, recovery after revert. Claude #89404 is the complementary negative flagship: the submitted suite is green while reviewer-supplied product-oracle evidence rejects the claimed-valid fixtures.

Evidence boundary

The receipt is explicitly REVIEWER_SUPPLIED_EXTERNAL_ORACLE. CounterProof does not claim it independently executed the Claude Code binary.