Add the conditional (status-aware) self-consistency rules deferred by #314

[NoopDog]

Follow-up to #314. The spike shipped only the flat/deterministic invariants (modality⟺assay, imaging exclusivity, sequencing-platform⟺histology, checksum/log inertness). The three status-aware rules from #314's rule set were deferred because they need engine work, not just YAML.

Rules to add

1. data_type ⟺ modality (Self-consistency check: cross-field semantic rules over each classified record (no ground truth needed) #314 rule 2): variants*/genotypes → genomic; expression_matrix/quantification → transcriptomic.*; peaks → epigenomic.*; images → imaging.histology; pangenome*/assembly*/sequence → genomic; array_signal → epigenomic.methylation.
2. reference applicability (rule 4): anchored types (alignments/variants*/annotations/peaks/signal) ⟹ reference_assembly must not be not_classified (a readable variants file with a blank reference is suspect; not_applicable is fine for de-novo). Unanchored types (reads/images/checksum/log/statistics) ⟹ reference not_applicable.
3. degenerate guard (rule 7): a readable data file with all five dimensions not_applicable is a red flag.

Engine work required (why this isn't just more YAML)

• A prefix matcher in the require clause (for "modality must be transcriptomic.*") — currently only when has prefix.
• A forbidden-state rule shape for rule 7: the when-state itself is the violation (there is no require that fails). The current model always needs a failing require.
• A "data file vs companion" predicate for rules 4 and 7 (a checksum with all-NA is fine; a BAM with all-NA is not). Needs a notion of "readable data file" beyond a single field.
• Extend the load-time matcher validation (_WHEN_MATCHERS/_REQUIRE_MATCHERS) to cover any new matcher forms.

Definition of done

• The three rules encoded and unit-tested (clean + violating records).
• Engine extensions above, with load-time validation kept in lockstep.
• Run over the corpus; triage any violations surfaced (these are more likely to fire than the flat rules).

Relates: #314. Note: index-inertness is tracked under #312, not here.