Thread surface: a subclass cannot super into the spawn (2 examples), Thread#backtrace of another thread missing (1) #700

Two independent Thread surface gaps, measured while fixing the blocking/GVL
defects in #698. Filed together because both are small and were found by the same
work; either can be closed on its own.

1. A Thread subclass cannot super into the spawn — 2 examples

Class.new(Thread) { def initialize; super { puts "ran" }; end }.new
# MRI : runs the block
# rbgo: ThreadError: must be called with a block

Isolated to Thread, not a general super-into-native problem: the same shape
into Array#initialize works on both engines. The cause is structural — rbgo puts
the spawn in Thread.new (a singleton method) where MRI puts it in
Thread#initialize, so a subclass overriding #initialize and calling super
reaches something that no longer carries the block.

Measurement: core/thread/new_spec.rb scores 6 against 8 for MRI 4.0.5 through our
own shim. Owning file internal/vm/thread.go. A fix means moving to
allocate-then-initialize, which is why #698 deliberately did not attempt it
alongside five behavioural fixes.

2. Thread#backtrace of another thread is unimplemented — 1 example

Costs core/conditionvariable/wait_spec.rb 6 against MRI's 7. Owning file
internal/vm/exception.go:522.

Why these are worth filing rather than leaving in a PR body

Both were named only in #698's report. A defect recorded in a PR body stops being
tracked the moment that PR merges, and three examples with a named owning file and
a one-line reproducer is exactly the shape that gets picked up by a later wave. The
same applies to two pre-existing sweep flakes measured in the same work, recorded
here so they are not re-diagnosed as regressions:

• core/kernel/require_spec.rb — 68 in 10/10 runs standalone on both engines; the
  67 appears only inside a 10-way parallel sweep.
• core/random/bytes_spec.rb — oscillates 8/9 on both engines (9 in 3/10 on main,
  7/10 on fix(vm): thread blocking and the GVL hand-off — five defects, one subject (#695) #698's branch, i.e. the same population).

That second one matters for the floor: it means a floor set with a margin of 0 or 1
over the measured total will go red on a flake. The ratchet's
UPDATE_FLOOR=1 writes the measured total exactly, so anyone using it should leave
a margin for this or fix the flake first.