Mechanical guard: no call entry may leave a param slot unbound (hazard class behind #1661) #1666

From the Opus 4.8 blind review of #1664, which PASSED. That PR closes the #1661 use-after-free at five call entries one by one, but it leaves the hazard CLASS open:

• env_new does not scrub recycled freelist slots;
• the JIT's OP_GET_LOCAL stays unchecked.

A future sixth call entry added without the null-fill would reopen the same use-after-free, and nothing mechanical would catch it. A runtime bounds check in the JIT is NOT the fix: GET_LOCAL is the hottest op (a call-heavy forced-JIT loop measured 477M Ir), and the check duplicates a structural invariant. Harden the invariant instead.

Done when

• A debug/ASan-build assertion at every frame push checks that every slot below local_count is bound (or null-filled), OR a source gate over env_new(... closure) / call-entry sites refuses one that does not null-fill. Either needs a planted sixth entry that turns it red.
• It is coordinated with Memory oracle has blind spots: freelists, arena, NaN-boxed slots and JIT code are invisible to ASan #1665 (trap tag / pool-off), which may make this redundant by construction.