Support the new ECOD distribution format - #1141

Fixes #1139.

ECOD redesigned its distribution at v294.1. Neither the version comment nor the column layout is recognised any more, so EcodInstallationTest.testVersion fails and — less visibly — anyone calling EcodFactory.getEcodDatabase() today gets an empty domain list with no exception thrown.

The version comment

parsedVersion stayed null and getVersion() fell back to the string that was requested, which for latest is the literal "latest" — the value the test asserts against. One pattern now matches both forms, so the historical releases keep working; a regex that only accepted the new form would silently break every pinned develop* version.

The columns

Rather than add a fourth positional special case, files that declare a column header — every release since develop101 — are now read by column name. That covers all four layouts at once and will not need revisiting the next time a column moves. Files older than develop101 have no header and keep the existing positional path, unchanged.

Everything else that moved with the columns:

• manual_rep holds True/False rather than MANUAL_REP/AUTO_NONREP;
• assembly_id is declared but empty on all 2,945,500 rows, meaning what NOT_DOMAIN_ASSEMBLY used to mean;
• the ligand list moved to ligand_comp_ids;
• ligand_pdbnum, the last column, is empty on 2,399,569 rows, so split("\t") was dropping it and making those rows look a column short. Trailing empty fields are now kept;
• f_name is empty rather than F_UNCLASSIFIED for unclassified domains — deliberately left as it is, see below;
• 1,573,776 rows (53%) describe domains found in AlphaFold models, with ids like P44140_F1_nD2 and an empty pdb column. EcodDomain is keyed by PdbId and cannot represent them, so they are counted and reported at INFO rather than each logged as an error.

Two failure modes that hid this are also closed: parsing nothing at all is now logged as an error instead of quietly returning an empty list, and the per-line NumberFormatException warning is capped like the three warnings beside it, rather than printing a stack trace per row.

getVersion() no longer parses the whole file

It reads the header. Parsing the current release in full to answer "which version?" costs 1,186 MB of heap and 9.8 s — and the test JVM is given -Xmx1500M by the root pom. The download is unchanged; only the parse is skipped. This is a small behaviour change, in that getVersion() no longer has the side effect of loading every domain, which the 7.3.0 target allows.

Measured

The first two are the counts EcodInstallationTest already asserts, so the old path is provably untouched.

Tests

EcodParserTest is new and entirely offline. EcodParser already accepts a Reader, so fixtures taken verbatim from the real files pin all five layouts, both version comment forms, a row ending in an empty column, an AlphaFold-derived row and a malformed row. 15 tests, no network, well under a second:

mvn test -pl biojava-structure -Dtest=EcodParserTest

That is the part I would most like kept. The distribution is 657 MB, so the only test that ever touched the current release was testVersion, asserting on one string; everything else pins develop204. That is exactly why a wholesale format change surfaced as a single one-line assertion failure, months after it happened.

testVersion now also parses the first few thousand lines of the local file — enough to notice a column change without building three million domains.

mvn test -pl biojava-integrationtest -Dtest=EcodInstallationTest

9 tests, 1 skipped (testAllVersions, already @Ignored for being slow), green in 124 s.

Note on CI

Branched off master, so its PR Build will still show the CATH failure until #1133 merges. That is #1138, not something new here.

Three judgement calls I would rather not make alone

@sbliven — you wrote this parser and know the format history far better than I do. If you have the time, I would value your eyes on it, particularly on:

1. unp_acc is read and discarded. develop291 carries a UniProt accession that EcodDomain has nowhere to put. Adding a field means touching equals/hashCode/toString/Comparable, so I left it out — but it is real data we are dropping.
2. An empty f_name is left empty rather than translated to F_UNCLASSIFIED. They used to be equivalent; now f_id classifies to a fourth level (1.1.1.3) while the name is blank, so I do not think they are the same thing any more. If you disagree it is a one-line change.
3. An empty assembly_id is mapped onto the old NOT_DOMAIN_ASSEMBLY semantics (assemblyId = uid). It is empty on every row of v294+, so assembly information is simply gone from the distribution; this keeps getAssemblyId() returning what callers expect rather than null.

Also deliberately out of scope, and worth a follow-up issue rather than growing this PR: valid_structure, ligand_binding, ligand_pdbnum, range_count and the five *_manual flags are new columns with no home in EcodDomain, and the AlphaFold half of the distribution needs a model that is not keyed by PdbId.

Unrelated but useful

ECOD now publishes ecod.latest.md5 in coreutils format alongside the distribution. That is a real hashURL for FileDownloadUtils.createValidationFiles(url, file, hashURL, Hash.MD5) — the path #1133 repairs and which, before it, could never write a hash file at all. Not wired up here; noting it so it is not lost.