`remeta gene`: bare std::invalid_argument: stod when a condition lead is out of range

[sklasfeld-p4]

Summary

• Component: remeta gene, conditional mode (--condition-list / --condition-htp)
• Version: v0.11.2-9462889 (source checkout: /home/sklasfeld/remeta, commit 9462889, matches the container banner)
• Severity: fatal — SIGABRT, exit 134

REMETA gene fails when conditioning is on and the condition list's first well-formed entry sits beyond the last record on that contig in some cohort's --condition-htp. Any earlier entry whose position has a record at or after it disarms the abort, so the same set of leads passes or fails depending on list order.

Call chain

Step	Location	What happens
1	src/run_htp.cpp:213	logs loading conditional variants...
2	src/run_htp.cpp:216	calls util::get_htp_variants(condition_list_file, condition_htp_files, chr)
3	src/util.cpp:79	per lead, per cohort: htp_reader.seek(cpra[0], pos)
4	src/io/bgz_reader.cpp:224-253	BgzReader::seek — the defect is here
5	src/util.cpp:81-82	while (rec.pos <= pos && !htp_reader.eof()) { rec = htp_reader.readrec(); ... }
6	src/io/bgz_reader.cpp:179-180	readline() returns this->seek_buffer, which is ""
7	src/io/htpv4_reader.cpp:205	readrec parses the empty line; stod("") throws

If the abort is in this path, src/run_htp.cpp:220 (found N variants to condition on in study i)
never prints. Presence or absence of those lines in <--out>.gene.log is the cheapest way to
localize the crash.

Root cause: check_eof() is called one line before seek_set is set

src/io/bgz_reader.cpp:239-253:

  int ret = tbx_bgzf_itr_next(this->bgzf, this->bgzf_tbx, this->bgzf_itr,
                              &this->buffer);
  if (ret == -1) {              // query matched nothing
    this->at_eof = true;        // line 242 — correct
    this->seek_buffer = "";     // line 243
  } else if (ret < -1) {
    throw runtime_error("tbx_bgzf_itr_next error");
  } else {
    stringstream ss(this->buffer.s);
    getline(ss, this->seek_buffer);
  }

  this->check_eof();            // line 251
  this->seek_set = true;        // line 252  <-- one line too late

src/io/bgz_reader.cpp:284-291:

void BgzReader::check_eof() {
  if (this->seek_set && this->seek_buffer == "") {     // line 285
    this->at_eof = true;
  } else if (!this->seek_set && bgzf_peek(this->bgzf) == -1) {
    this->at_eof = true;
  } else {
    this->at_eof = false;                              // line 290
  }
}

On a reader's first seek seek_set is still false, so the guard at line 285 cannot fire.
Control reaches the bgzf_peek branch, and unless the underlying stream is physically at EOF,
line 290 resets at_eof back to false — discarding the correct value just set at line 242.

Consequences, in order:

1. HTPv4Reader::eof() (src/io/htpv4_reader.hpp) reports "not at end".
2. src/util.cpp:81 enters its loop and calls readrec().
3. readrec → BgzReader::readline (src/io/bgz_reader.cpp:172). seek_set is true by now, so
   line 180 takes line = this->seek_buffer — the empty string set at line 243.
4. readrec parses "". Every >> extraction fails, leaving each token empty.

Trigger conditions

The abort requires a tabix query that matches nothing. src/io/bgz_reader.cpp:248 queries
tbx_itr_queryi(tbx, tid, position - 1, HTS_POS_MAX) — i.e. [pos, end-of-contig], open ended.

• It matches whenever any record exists at or after pos on that contig. Proximity to the
  nearest record is irrelevant: a lead in a gene desert lands on the next record and is handled
  correctly.
• It matches nothing only when pos > (last record's position on that contig in that file).
• A contig missing from the index fails earlier and differently: tbx_name2id < 0 →
  runtime_error("called seek with invalid contig") (src/io/bgz_reader.cpp:233-236)

Only the first seek on a given reader can trigger it. From the second seek onward seek_set is
already true, the guard at line 285 fires, at_eof stays true, and src/util.cpp:81 skips the
loop — remeta simply does not find that lead. get_htp_variants builds one reader per cohort
(src/util.cpp:69), so in practice it is the first well-formed variant ID in the condition list
that decides whether the run aborts. Malformed IDs are skipped at src/util.cpp:73-76 and do not
count.

Confirmed against v0.11.2-9462889 on a two-cohort fixture with two leads, both absent from the
HTP, differing only in position — 1:60000 before every record, 1:140000 after the last one:

condition list	result
1:60000 alone	PASS — and silently not conditioned on
1:140000 alone	ABORT, what(): stod
1:60000, then 1:140000	PASS — the fatal lead is disarmed
1:140000, then 1:60000	ABORT — the same two leads, only the order changed

Two consequences worth stating plainly:

• Survival depends on clumping output order. The same data can abort or not between runs if
  that order is not stable.
• The non-fatal case is not harmless either. A lead whose seek matches a record that is not
  the lead is silently not conditioned on — no warning, and the run degrades to marginal for that
  lead.

The practical scenario. Condition leads come from LD clumping against a common/imputed
reference panel, while --htp for the gene tests is an exome/PAV HTP. Exome variants stop well
before the end of a chromosome, so a common lead beyond the last PAV row matches nothing. The
effective ceiling is the minimum over cohorts of the last row's position, since any one
cohort's empty query is enough.