feat(apollo): rebuild the Apollo release package as a tested CLI

[jbrestel]

Replaces the hand-maintained script that built the Apollo release package. That
script kept its thresholds, argument handling and rename detection in a file in
a home directory, so nothing ever exercised them — it produced an empty release
for build 71 and exited 0.

Model/bin/createApolloReleasePackage is now thin wiring over
ApiCommonModel::Model::ApolloRelease::*, where every rule with a wrong answer
worth catching is a testable class method.

Two phases

• --report — minutes, changes nothing. Portal + Apollo + overlay + rename
  resolution + reconciliation, printed. This is what goes to the curation team.
• --generate — hours. Builds the package for the approved roster and writes
  the command files a human then runs against Apollo.

It never calls a mutating Apollo endpoint. The step between "computed" and
"applied" is always a file a person reads.

Modules

Portal	the UniDB organism list, normalised; carries both abbrevs
Apollo	the live roster — the SEED for the release
Overlay	the human add/remove decisions, reason mandatory
Reconcile	set logic → update / add / prune / rename / exception
Rename	assembly identity from the .fai, for what the database cannot explain
Absolutize	site-relative → absolute URLs, with a mandatory post-condition
Generate	the per-organism package
Commands	the curl and groovy command files
Report	the artifact the curation team approves from
Cli	options, preflight, sanity gates, rename resolution

Verified against production

A real --report reconciled the whole roster exactly: 453 update / 34 add /
1 prune / 1 rename / 4 exception. It detected the cneoJEC21 → cdenJEC21 rename
from the database with its 14 annotations preserved, and declined the
cglaCBS138 → nglaCBS138 merge — nglaCBS138 is already in Apollo, and
performing that rename would leave two Apollo organisms sharing one directory,
which Apollo.pm refuses to load on the next release.

508 tests across 11 files.

The failure modes this is built around

• An add plus a prune for the same genome — creates an empty Apollo organism
  and orphans the curation work, from the same API and the same inputs as the
  correct behaviour. Reconcile::assertInvariants guards it and is deliberately
  noisy.
• An update silently re-publishing hidden organisms — the previous script
  hardcoded publicMode: true. Visibility is now echoed back from Apollo, and
  absent it, the emitter dies rather than defaulting.
• An empty release that exits 0 — assertPortalSane, assertApolloSane,
  assertUpdateBucketSane and assertRosterNonEmpty. The last one takes no
  force argument, deliberately: --force accepts omitted candidates and an
  empty update bucket, but never licenses a package with nothing in it.
• A silently missed URL rewrite — a track that 404s inside Apollo loads,
  appears, and is empty. Absolutize::assertNoRelative is the post-condition
  for rewrite, sharing one pattern so the two cannot drift.

Known open, not addressed here

From a CLI review, all pre-existing and none blocking a run:

• A rename bypasses Portal::qualifies, so an organism demoted this build would
  still be generated and repointed rather than surfacing as an exception.
• report.txt / report.tsv are written after generation, so an hours-long
  run that dies late leaves no paper trail.
• narrowResult's prune filter is unreachable from the CLI — generationRoster
  rejects a prune abbrev first — so a partial --generate always emits zero
  approved prunes.
• PORTAL_FLOOR makes --project unusable for non-UniDB models.
• _write uses >:raw, so a non-ASCII organism name lands as invalid UTF-8 in
  the archival report.

Note for the reviewer

e8670e13a ("fix simple error in TranscriptsFromGenes SQL", one line in
geneQueries.xml) is the base commit of this branch and is unrelated to Apollo.
It rides along because the branch was cut on top of it. Say if you want it
split out.

🤖 Generated with Claude Code