Untitled

Day 26 evidence

Date: 2026-09-20

Source baseline

• Implementation commit: aacc1a350cb32eeb647669770112568e8e81f25b
• Text kernel SHA-256: 1c6c5d3f5858f1c7e627d7762e596e8c8e757f23d15aad81c6f583f82a06d8f1
• Graphics text kernel SHA-256: 95d3b1bf57ddf620a25db0681c409337506e55366bf07b51eafe6f47c2758008
• Graphics text disk image SHA-256: a73be1c6290e3cc2b7571365ffaa089602c6ee6d85d067b37f7a31ccfc87b5b6
• Screenshot SHA-256: 90e06a60579f6546ef7d0d6127dbdd810a906d04021ff1ef5e71d3bf98c91892

The implementation adds a separate mbr-graphics-text.img. Historical text
images remain on BootInfo v1, and the Day 25 graphics image remains a separate
BootInfo v2 regression target.

Font provenance

The committed table keeps printable U+0020 through U+007E from
dhepper/font8x8 font8x8_basic.h, pinned to commit
8e279d2d864e79128e96188a6b9526cfa3fbfef9. The upstream README describes the
font as Public Domain. third_party/font8x8-NOTICE.txt preserves that statement,
the author, URL and commit. This record does not make an independent legal
determination. Bytes outside printable ASCII are replaced with ?; this is not
UTF-8 or Chinese text support.

Environment

• QEMU 8.2.2 (Debian 1:8.2.2+ds-0ubuntu1.18)
• SeaBIOS 1.16.3 package image
• GDB 15.1
• GCC 13.3.0 and GNU Binutils 2.42
• NASM 2.16.01
• Machine: pc-i440fx-7.2, TCG, qemu32, 1 CPU, 64 MiB

The unpacked toolchain under /tmp/build-an-os-tools is an execution dependency
of this environment and is not included in the attachment.

Specialized command

PATH=/tmp/build-an-os-tools/bin:$PATH \
LD_LIBRARY_PATH=/tmp/build-an-os-tools/root2/usr/lib/x86_64-linux-gnu:\
/tmp/build-an-os-tools/root2/lib/x86_64-linux-gnu \
QEMU_MODULE_DIR=/tmp/build-an-os-tools/root2/usr/lib/x86_64-linux-gnu/qemu \
make check-graphics-text \
  QEMU=/tmp/build-an-os-tools/bin/qemu-system-i386 \
  GDB=/tmp/build-an-os-tools/bin/gdb-multiarch \
  BIOS=/tmp/build-an-os-tools/root2/usr/share/seabios/bios-256k.bin \
  QEMU_DATA=/tmp/build-an-os-tools/root2/usr/share/qemu

Exit status: 0.

Runtime observations

The graphics text image reached the Day 26 completion breakpoint with these
values:

STATE pages=16 rollback=7 free=16046/16030/16046 clip=2/2 pixels=204 guard=0 wrap=1 scroll=8 cursor=6,10 hash=0xd84d1c49
PASS: paged backbuffer, rollback, clipping, glyph, terminal and flush verified

The serial log independently reports:

D25 VIDEO MODE 13h
D04 PM OK - BootInfo v2
D26 BACKBUFFER pages=16 rollback=7 free=16046/16030/16046
D26 CLIP clipped=2 rejected=2 pixels=204 guard=0
D26 TEXT wraps=1 scrolls=8 cursor=6,10 A=c,1e,33,33,3f,33,33,0
D26 FRAME hash=d84d1c49
GRAPHICS TEXT OK clipped, rendered, flushed and freed

The free-page sequence proves two ownership paths. A forced failure before the
eighth allocation returned all seven acquired pages (16046 before and after).
The normal surface then held 16 pages (16030) and returned all of them after
the framebuffer copy (16046).

The clip probe called the same rectangle primitive with INT32_MIN, dimensions
of UINT32_MAX, a rectangle wholly outside the display, and an 8x8 rectangle
crossing the bottom-right corner. The two intersecting calls wrote exactly 200
and 4 pixels. The unused 1,536 bytes at the end of the 16-page surface retained
the 0xA5 guard value.

GDB read all 64,000 framebuffer bytes and recomputed FNV-1a 0xd84d1c49, equal
to the hash computed over the paged surface before the copy. It also checked the
eight rows of A (0c 1e 33 33 3f 33 33 00) and corresponding foreground and
background pixels. The QEMU screenshot is 640x400 because the 320x200 legacy
scanout is doubled by screendump.

Loader bounds

kernel-graphics-text.bin is 92,128 bytes, below the 128 KiB file limit. Its
ELF program headers report a 92,096-byte executable segment and a data segment
with 37,424 bytes of memory; the complete runtime image remains below the
existing 256 KiB limit. The 64,000-byte frame is not static BSS: it is allocated
as 16 physical pages at runtime.

Cumulative regression

After the specialized check, the following targets all exited zero:

check-graphics
check-pm
check-pipe
check-shell
check-fd
check-fat
check-block
check-process
check-syscall
check-elf
check-memory
check-runtime-limit
check-keyboard

The combined log is docs/day26-cumulative.txt. It demonstrates that the new
translation unit did not change Day 25 framebuffer bytes or the older text,
storage, process and user-mode paths.

Frozen attachment verification

• os-day-26.zip SHA-256: 32addbe3f40b5435114d9d0c233b40f760e61534320712f1f59d044762713ad2
• The archive contains no build/ directory.
• A new temporary directory extracted the archive and ran build,
  check-graphics-text, check-graphics, and the thirteen cumulative targets.
  Every target exited zero.
• The rebuilt text kernel, graphics text kernel and graphics text disk image
  matched the pre-freeze SHA-256 values byte for byte.

The post-freeze receipt is docs/day26-zipcheck.txt. It is intentionally not
inserted back into the ZIP, because doing so would change the artifact whose
hash the receipt records.

Evidence boundary

• The surface gives drawing code a stable off-screen frame. Copying 64,000 bytes
  to VGA memory is not synchronized with vertical retrace, so the experiment
  does not prove tear-free presentation.
• The page vector removes the need for physically contiguous frames, but it is
  not a contiguous virtual mapping. Every byte access performs a page lookup.
• The terminal is 40x25 fixed cells with one 8x8 font and indexed colors. It has
  no Unicode shaping, font fallback, variable width, scrollback history or
  user-mode terminal protocol.
• The screenshot proves the QEMU display consumed the bytes. Bounds and rollback
  claims come from guest assertions and GDB state, not from the screenshot.