[Ehsan-hash]

Fixes #405.

csv.DictReader silently keeps the last value for a repeated header. In paired statistics this could reverse the reported comparison. Reject duplicate decoded column names before reading observations, running inference, or writing results, matching the existing paired-verification boundary.

Validation:

• 12 focused checks passed for path/stream callers, duplicate score/unit/metadata/quoted headers, output preservation, ordinary numerical controls, and the actual API queue plus direct worker execution.
• 54 affected handoff and evidence checks passed.
• The pristine implementation analyzed the reported duplicate input as improvement −90; the unique-column control gives +1. The original-source regression run has 6 expected failures and 5 passing controls. One duplicate-unit case already rejected for too few distinct units; it now rejects the ambiguous header earlier.

The worker checks use independent SQLite/data/temp roots. Ambiguous input retains its exact source bytes but publishes no metrics or Analysis; the valid control retains the expected result and provenance. These checks exercise the direct worker branch, not daemon lifecycle transitions. No live provider, research experiment, PostgreSQL, browser, or full-suite validation is claimed.

[Ehsan-hash]

Thanks for merging these fixes, @mikamikasuki!