valgrind (3.19.0) locally on macOS docker container seems to detect this leak but not CI.
$ cat modulemeta.test
modulemeta | .deps |= length
"c"
{"whatever":null,"deps":6}
$ valgrind --error-exitcode=1 --leak-check=full -q ./jq --run-tests < modulemeta.test
Test #1: 'modulemeta | .deps |= length' at line number 1
*** Insufficient results for test at line number 3: modulemeta | .deps |= length
0 of 1 tests passed (0 malformed, 0 skipped)
==14129== 482 (392 direct, 90 indirect) bytes in 1 blocks are definitely lost in loss record 7 of 8
==14129== at 0x48407B4: malloc (vg_replace_malloc.c:381)
==14129== by 0x13BC78: jv_mem_alloc (jv_alloc.c:122)
==14129== by 0x1371C6: jvp_object_new (jv.c:1521)
==14129== by 0x145840: parse_token (jv_parse.c:167)
==14129== by 0x145840: scan (jv_parse.c:684)
==14129== by 0x145840: jv_parser_next (jv_parse.c:793)
==14129== by 0x1463CB: jv_parse_sized (jv_parse.c:865)
==14129== by 0x136BD9: run_jq_tests (jq_test.c:183)
==14129== by 0x1370FA: jq_testsuite (jq_test.c:34)
==14129== by 0x13135C: main (main.c:582)
==14129==
Extend jv_number to use decNumber for storing number literals. Any math
operations on the numbers will truncate them to double precision.
Comparisons when both numbers are literal numbers will compare them
without truncation.
Delay conversion of numbers to doubles until a math operation is performed,
to preserve precision. A literal jv_number will only need conversion to
double once, and will reuse the resultant double on subsequent
conversions.
Outputting literal jv_numbers preserves the original precision.
Add strong pthread requirement to manage contexts/allocations for
converting numbers between their decNumber, string, and double formats.
