Warn early when project metadata implies multiple samples but the data is pooled-only (no barcodes to separate them)

### Observed
Project context listed multiple samples/conditions, but the raw paired-end FASTQs carried no per-sample barcode/index, so the samples could not actually be demultiplexed. The agent eventually concluded that only pooled analysis was valid and sample-level comparisons were impossible -- but only late in the workflow, after sample-level labels had already propagated through the analysis context.

### Expected
When the project/user context implies N samples but the observed file structure lacks any barcode/index to separate them, Orbit should warn early and plainly: the data can only be analyzed as pooled unless the user provides demultiplexing information -- before building sample-level scripts.

### Relationship to #288
Sibling, not a duplicate. #288 validates *unique sample labels* before launching DESeq2 (a labels-collision check in the DESeq2 path). This is the upstream, more general check: *can the inputs even be demultiplexed to the sample granularity the metadata claims?* It fires earlier (at metadata/file-structure reconciliation) and isn't DESeq2-specific. Could be folded into #288 as the same guardrail family if maintainers prefer.

### Environment
Orbit 0.4.1, Linux x64, provider anthropic / claude-sonnet-4-6.

Start at the metadata/file-structure reconciliation step and trace how sample-level labels reach script construction. Compare that guardrail with the DESeq2 label check described in #288. Done means pooled-only FASTQs trigger an early, plain warning that demultiplexing information is required for sample-level analysis.

Written by the indexing model from the issue text.