fix: verify reconstructed sequences by genome name

[mmolari]

Unifies sequence verification across build, merge and reconstruct behind one name-keyed helper. Covers §4.4 of dev/design/merge.md.

Matching by name is the only option left once merge renumbers path ids: neither record order nor FastaRecord::index can pair genomes up. The §4.3 checks from the previous PR make names a valid key, since they are now guaranteed unique.

This turned out to fix two latent bugs rather than just unify style:

• build --verify panicked. It indexed &fastas[actual.index], valid only when the input records' indices are exactly 0..n-1. Any programmatically assembled record set breaks that — including this repo's own test helper, whose right half carries indices [3, 4, 5] for a 3-record slice. Every existing test escaped it by passing verify=false, so build(.., verify=true) had no coverage at all.
• reconstruct --verify was unreliable. Records were paired positionally, so a verification file with surplus records silently passed (exit 0); too few gave a misleading expected length 0 but got N; and a merged graph failed spuriously with expected length N but got N at identical lengths, because whole FastaRecords were compared and index differs after relabeling.

pangraph/reconstruct.rs is a new home for reconstruction (it is a graph operation — pangraph/pangraph.rs was reaching up into commands::) and holds the shared API: path_ids_by_name, reconstruct_genome, reconstruct_by_name, sequences_by_name, verify_genome, verify_graph_sequences. compare_sequences is deleted.

verify_graph_sequences takes a GenomeCoverage argument, which was not in the original §4.4 sketch: build's intermediate clade graphs hold only their own clade's genomes and need Partial, while the final build graph, a merged graph and reconstruct --verify require Complete. Making the final build check Complete is a small gain — nothing previously noticed a genome going missing.

reconstruct --verify keeps streaming rather than adopting the map: it is the only one of the three that does not otherwise need every genome resident, and a map would make peak memory proportional to total genome length (~2.5 GB for 500 E. coli genomes, vs ~10 MB today). It shares verify_genome and its error formatting.

Errors now name the genome and, for equal-length differences, the first differing position with context:

Sequence mismatch for genome 'NZ_CP027173.1' at position 100 (length 132518):
  expected: GAACCCGAACAATCCGTTGTA
  actual:   GAACCCGAACTATCCGTTGTA

13 unit tests and 7 new integration tests, including the headline case (verify a merged graph against its genomes in reversed order) and an explicit regression test for the build --verify panic. docs/docs/reference.md regenerated for the check-cli-docs CI gate.

Also fixes a pre-existing break caught by cargo clippy --release: the merge integration tests called sanity_check, which is #[cfg(any(test, debug_assertions))] and so does not exist for integration tests in release mode.