#520 (comment) mentions a second issue:

2. This write call crashes. I'm not sure why. This only seems to be happening on nightly-2025-11-21 and later. It works on nightly-2025-11-20. AFAICT this isn't caused by the fix for the first issue. Shrinking the file size by other means still has this line failing, so I don't think it's the opt-level. Removing write call allows the kernel to boot.

Let's see whether the CI passes.

All the failed relocations are from the .start section, referencing the .bss section.

Where did you see that? I only see references from .text e.g.

            rust-lld: error: /home/freax13/code/bootloader/target/i386-code16-stage-2/stage-2/deps/bootloader_x86_64_bios_stage_2-836cb229c0ae0c00.bootloader_x86_64_bios_stage_2.23a636a39f56531e-cgu.0.rcgu.o:(function bootloader_x86_64_bios_stage_2::start: .text._RNvCs33L8l6sfb8U_30bootloader_x86_64_bios_stage_25start+0x25a): relocation R_386_16 out of range: 74745 is not in [-32768, 65535]; references section '.bss._RNvNvXNtCs33L8l6sfb8U_30bootloader_x86_64_bios_stage_24diskNtB4_10DiskAccessNtB4_4Read10read_exact7TMP_BUF'
            >>> referenced by bootloader_x86_64_bios_stage_2.23a636a39f56531e-cgu.0

The "Build examples" job is failing because it's depending on the bootloader crates from crates.io instead of using a path dependency. We can ignore this for this PR.

Where did you see that? I only see references from .text e.g.

Yep, you're right! I blindly assumed that the start function in inlined into _start and thus ends up in the .start section. But it looks like this does not happen.

The "Build examples" job is failing because it's depending on the bootloader crates from crates.io instead of using a path dependency. We can ignore this for this PR.

Yes, but the Test jobs time out too. I can reproduce this locally, it seems to crash/hang during parsing a RawDirectoryEntry. During my println debugging I noticed that crate::screen::Writer.write_str works fine, but writeln!(crate::screen::Writer, ...) hangs as soon as there is at least one fmt argument.

Without this PR, it works as expected...

I opened #522 with alternative approach (keep the linker script unchanged and switch back to legacy mangling).

Closing in favor of #522 for now.