Nothing executes the code a spec generates, so correctness is unmeasured #3903

Nothing in the pipeline has ever executed the code a spec generates.

A spec is accepted when a bee asserts its acceptance criteria are met and the
Queen agrees. Every criterion the review could judge mechanically — t27c parse,
parse-complete, typecheck, empty-body count — asks whether the spec PARSES.
None asks whether what it generates WORKS. queen food counts empty function
bodies and reports the number as progress, which scores a body full of wrong
code above an empty one.

What that has cost

Measured 2026-09-16 by generating all 946 specs to Zig and running zig test:

• 541 pass — compile and pass their own tests
• 384 do not compile at all
• 21 compile, run, and hit @panic("not yet implemented")

That distribution was unknown and unmeasurable before the measurement existed.

Acceptance criteria

• tools/oracle/run.sh generates every spec into a tree mirroring specs/ and
  runs each under zig test
• The oracle reports the file each error is IN, not the file under test
• tools/oracle/baseline.tsv records the passing set and the run fails when a
  spec that used to pass stops passing
• The run does not fail merely because a spec has never passed

No ## Boundary on purpose: the work is already done in #3882. A boundary here would let a bee claim it and reimplement what is already in review.