Repro scripts for findings 19, 20, 32, 33, 42; finding 32 corrected

[zmaril]

Self-contained reproduction scripts in docs/hunt/repro/. Each one:

• creates a minimal Cargo project in a temporary directory, writing every file with a heredoc;
• runs the exact commands (cargo +<toolchain> …, plus clang for the C side of the ABI cases);
• checks the output and ends with REPRODUCED or NOT REPRODUCED.

Each script's header says how mirth found the bug. None came from a public crate.

Script	Finding	Severity	Run here
19-riscv-stack-arg-extension.sh	riscv64: an i32 passed on the stack to extern "C" isn't sign-extended (clang marks it signext)	high (stable, tier 2, silent wrong values across FFI)	REPRODUCED on stable and 1.80.0
20-riscv-float-pointer-struct.sh	riscv64: struct { float; void* } goes in a float and an integer register; clang uses two integer registers	high (same)	REPRODUCED on stable and 1.80.0
33-rustdoc-empty-nested-use.sh	cargo doc panics on use {{}};, which cargo build accepts; a nightly regression since 2026-09-26	medium	REPRODUCED (09-25 is fine, 10-06 panics)
32-iter-chain-solver-regression.sh	rejecting an iterator chain over the recursion limit takes about 8× the CPU and gives 75 errors instead of one under the new solver	low	REPRODUCED
42-dead-code-used-trait.sh	dead_code calls two traits "never used", and deleting them breaks the build	low (stable, warn-by-default)	REPRODUCED on stable and 1.80.0

Correction to finding 32. It was first recorded as a medium compile-time regression on valid code. That was wrong: the 200-map program is over the recursion limit and every toolchain rejects it, and the original timings never checked the exit status. Under the limit (126 maps), the new solver isn't slower; it's faster. This PR rewrites the facts doc and the hunt.md row to say what does differ: how slowly and noisily the program is rejected. Severity is now low.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QXiEXbESemwqMLYKaWLDbT