ghostty-rs/docs/rewrite/notes/review-rv12-codex.md

Codex final verification (RV12, 2026-10-12)

All three RV11 reproductions are fixed at 4bbe0d7. Exact parity remains incomplete: 2 newly verified diagnostic gap groups and 1 known visual gap—0 P1, 0 P2, 3 P3. No files were modified; the worktree remains clean.

RV11 finding	Verification
Duplicate benchmark enum arguments	Resolved. Earlier invalid values are rejected, including before empty resets. Sixteen differential checks matched Zig.
Interior-NUL theme paths	Resolved. Absolute and named paths load their prefix, producing #262427 without diagnostics in both implementations.
Filesystem diagnostic descriptions	Original Linux cases resolved, including absolute and relative NotDir paths. The replacement helper has the portability defect below.

Concrete remaining gaps:

1. P3 — Platform error codes interpreted as Linux errno values. zig_error.rs:37 unconditionally applies the Linux table. Windows access-denied code 5 becomes error.Unexpected; Darwin ELOOP=62 and ENAMETOOLONG=63 also become Unexpected. Verified statically against platform definitions and call sites; native execution remains untested.

2. P3 — Incorrect diagnostics for dot theme names. theme.rs:187 uses Rust’s file_name(). Live differential checks show theme=. and theme=.. report “cannot include path separators,” whereas Zig reports “it is a directory.” Both reject the theme.

3. P3 — Known sprite stroke rasterization differences. canvas.rs:489 retains tiny-skia stroke geometry. Recomputing the saved VP captures confirmed 196 differing pixels per page scenario, maximum channel delta 86; tabs/splits and chrome are identical. This is the documented curly-underline/diagonal-glyph difference.

The fresh sweep covered the 461-file inventory: 428 non-build files and 33 build files, with counterpart, public-function/export, stub and targeted behavioral checks across all subsystems. No wholly absent feature was established. This is breadth coverage, not proof of exhaustive behavioral equivalence.

Validation produced:

• 2,759 existing-binary tests passed, four ignored; four encountered read-only filesystem restrictions.
• 6,300 configuration diagnostic comparisons, covering all 202 fields and compatibility handlers: only the dot-theme differences above.
• 2,700 configuration value/diagnostic comparisons: only expected debug-default differences.
• Matching actions, default keybindings, theme/font listings, zsh/fish completions; WASM smoke test passed.
• The documented 390,000-case terminal campaign remains supporting evidence; it was not rerun. Genuine Zig memory bugs/UB were excluded.

The benign +list-colors difference also persists: Rust flushes all 15,434 bytes, while Zig stops at 12,297. It is excluded from defect counts.

Inherently unverifiable natively here: Apple Metal/IOSurface/CoreText, AppKit/UIKit/Carbon integration, Mach crash handling and Apple framework packaging; Windows ConPTY/CreateProcessW, native filesystem/environment behavior and exception handling.

Fresh Linux GPU/compositor, portal/D-Bus and interactive GTK runs were not performed. Those are runtime coverage limitations, not inherently untestable Linux functionality.