Fix critical correctness bugs in SASS rewriter and interceptor

[Zorg33]

Summary

• ELF header parsing: e_phnum/e_shnum were read from offset 48/50 (inside e_flags field) instead of the correct 56/60, causing the rewriter to process zero sections and produce unmodified output
• Multi-kernel ELF splice: cum_growth double-counted offsets already shifted by the inner loop, corrupting binaries with 2+ .text sections
• Predicated FFMA split: FMUL was forced to always-execute (0xFF), clobbering r_dest when the predicate is false and temp_reg == r_dest
• Opcode matching: "FMA" in op incorrectly matched DFMA (double) and HFMA2 (half), splitting them with float FMUL/FADD opcodes
• Jump table safety: verified_jump_tables was computed via BRX/LDC tracing but never consulted — .nv.constant* values matching PCs were blindly patched, risking corruption of float/integer literals
• Per-kernel REGCOUNT: global max_reg set every kernel's register count to the max across all kernels, reducing occupancy
• Program headers: p_filesz/p_memsz/p_offset were not updated after .text expansion
• Interceptor get_elf_size: computed size from header positions + 8192 padding instead of walking section data extents
• Thread safety: declared g_mutex was never locked in patch_and_load
• Minor: write() return check, file-open guards in tests, R255 filtering, symtab bounds checks, SHT_DYNSYM support, unused deps removed

Builds on c500a4c (Fix elf processing) with additional correctness fixes.

Test plan

• Build with ./build.sh on sm_70 hardware
• Run ./run_test.sh — all 4 suites (basic, comprehensive, stress, runtime JIT) should pass
• Verify rewriter actually transforms FFMAs: check [Rewriter] Patch Complete output shows Sections: N where N > 0
• Test with a multi-kernel cubin (comprehensive test) to validate the splice fix
• Test with predicated FFMA kernels to validate predicate preservation

🤖 Generated with Claude Code