Simplification round: collapse five repeated shapes, and fix the two cache leaks one of them bred

[JamesKane]

A reuse-and-concision pass over the workspace, looking for shapes that had been copied rather than
named. Net −1,180 lines across five consolidations, plus one behaviour change that falls out of
the third.

Sequel to the 2026-06 round, which split the two God files. That one moved code between files; this
one removes copies of the same idea.

The behaviour change (fix(store):)

consensus_archaic, consensus_roh, consensus_painting and consensus_archaic_segments were the
same module four times over — same struct, same upsert/get/delete, same test — with a table name and
three column names swapped. ~420 lines to say one thing.

The copies had already cost something. Both places that purge a subject's derived results enumerate
the cache tables by hand, and each had fallen behind the set by a different amount:

• biosample::clear_data — the "reset this subject's analysis" path — only ever named
  consensus_painting. ROH and both archaic caches survived a clear, still keyed to a consensus
  signature that no longer existed.
• purge_alignment_derived named three of the four and left the Tier B archaic segments behind,
  keyed to an alignment that had just been deleted.

Both now iterate sig_cache::ALL. That is the fix; the deduplication is what makes the list exist.

The schema is untouched — the columns really are named differently per table for historical
reasons, so each cache carries its own names and get aliases them back to a common shape. Table
names are &'static str constants, never caller input. The tests run over ALL, including one the
four copies could not have written: that writing one cache does not disturb another sharing a column
name (segments belongs to both consensus_painting and consensus_archaic_segments).

The rest (behaviour-preserving)

• navigator-sync — push_create_inner/push_put/push_delete/pull_list each carried their
  own copy of refresh-on-401 + exponential backoff + offline flag. A change to the retry policy had
  four places to land, on the path where getting it wrong means hammering a PDS or dropping a
  publish. One with_resilience driver; the four methods become one line each.
• navigator-app — three clients reached the AppView independently and converged on the same two
  request shapes. social_post and exchange_post were byte-for-byte identical (a doc comment even
  said one mirrored the other), social_get/exchange_get_poll differed by a closure, and
  sync.rs/matching.rs open-coded it twice more. New appview.rs owns the URL, both shapes, the
  transport-error mapping (16 sites) and the non-2xx classification.
• navigator-ui/cli.rs (2,415 → ~2,000) — ~70 four-line match blocks were standing in for ?
  in functions that return an exit code. An ExitCode trait lets a single cli_try! arm cover both
  error types; require_subject takes the twelve copies of the not-found block.
• navigator-domain — ::new constructors for the workspace aggregates, and 91 exhaustive
  literals rewritten onto them. Fields that carry intent now lead, defaults sit behind a ...
  Deliberately not applied to the three store row→domain mappers, where the exhaustive literal is
  the safety property.
• navigator-ui/chrome.rs — alignment_has_bam (×5), cancel_button (×4), danger_button (×5).
  The cancel button's non-obvious detail — it disables itself once clicked, because cancellation is
  cooperative — was documented in one copy and absent from the other three.

Left alone on purpose

The 4× gVCF read_line preamble: an iterator version trades a reused line buffer for a per-line
allocation in a read path, and a closure version turns every continue into return Ok(()) through
60–100-line bodies. And the three egui two-column row closures: 26 call sites churned to save 9
lines.

Verification

cargo test --workspace	0 failures
cargo clippy --all-targets --workspace -- -D warnings	exit 0
cargo fmt --all --check	clean
all six commits cargo check --workspace --all-targets	PASS

The history is bisectable — each commit was built individually, not just assumed to compile. The
first sweep caught a real failure (haplogroup.rs spans two themes and used appview_url a commit
before it existed); fixed by reordering the last two commits, with git diff confirming the final
tree stayed byte-identical to the one the test suite ran against.

🤖 Generated with Claude Code