[v0.4] Plant evidence states what it shows: generator refusals by locus, SP-16's forged-program plant, the SM-32 records (follow-up to #167) #184

Follow-up to #167 (reviews of #157, #163 and #151).

1. assert_targeted (crates/conformance/src/cases/preservation.rs:91-101) counts any generator refusal as a caught lowering plant. It checks only that the function owns a relation, not that the refusal names the mutated function and site. The SP-15 record (VERIFICATION.md:1471-1530) shows that the six observed refusals do name the site, but the case does not assert it. Both Buffer plants are refused by the generator, so no Buffer mutation reaches the kernel.
2. SP-16's plant "forged program literal" forges the crate term, not the program row.
3. VERIFICATION.md:8870-8885: the two SM-32 plants cite bit_contribution, is_fixed_integer() and is_unsigned_fixed_integer(), which exist in no version of the code, and an output ("check fails with LLT4001: ()") the harness does not print.

Definition of done:

• A refusal counts as caught only when its message names the mutated function and site, and the case asserts it.
• At least one Buffer mutation is carried to the kernel by a plant that bypasses the lockstep refusal, or the SP-02 statement says which kinds the generator alone refuses.
• SP-16 gains a plant on the program row, or the plant is renamed.
• Both SM-32 plants are re-run against widen.rs and semantic.rs, with the real edit and the real output recorded.
• just vv green.