Step 12b: two ratified defaults, undone
Two runtime choices a worker made and Fable ratified are being reversed, because each lets a program pass while quietly not meaning what the spec says. Small step, two parts.
Part A: memoization leaves the reference semantics
Step 7 added a cache of pure-call results under mo run. Remove it from the interpreter: the interpreter is the executable reference semantics (chapter 7) and must run every body every time, checking every contract every time. Keep the profiler’s finding as a note in the code: speed comes from the C backend (step 13), not from the reference. Re-record the logstat-4k and kv-10k-get bench rows and put before and after in the commit message; a slowdown is expected and accepted.
Part B: every never runs on every test
A never block is evaluated at the end of every test, test rejects, and property run, over the values that run produced, not only under --sim. Recording covers structs, enums, and primitives held in any binding, field, message, or state (extend the recorder; the cost is accepted in mo test). A never over a type the recorder cannot cover is a compile error, MO0324, “this never cannot be checked” (chapter 2, contract laws), never a silent skip. The corpus gains examples/contracts/never-trips.mo: a never that is false on plain test data, with a test rejects expecting it, so mo test without --sim proves the check runs. verified: contracts on the line now means never was checked.
Done when
zig build test green, the bench rows re-recorded, the new corpus file trips without --sim, pushed, decisions listed. Fable updates the decision log rows to overturned.