Finding-by-finding (framing revision 11; bites via scripts/bite
against 6793edc):
1. Style-span coordinate translation belongs to the BUFFER. A new
BufferStyleSpanTranslator is attached by
pmacs.buffer.add_style_overlay and sees every edit exactly once —
bypass writes, undo/redo, remote CRDT ops — independent of window
count or visibility; the window-attached BufferStyleOverlay
copies are render-only (on_edit removed). Pre-fix each attached
view translated the shared store: start_run's explicit attach
duplicated the after-switch hook's (switch_buffer fires it
synchronously), so the normal path shifted later spans TWICE per
byte-delta rewrite, splits multiplied further, and a hidden
buffer shifted ZERO times. The redundant attach is removed;
correctness no longer depends on attachment discipline. Bites:
per-cell rendered assertions active (red a, blue bc, CR, red é →
é red, b/c blue) and hidden (run finishes with the buffer in no
window; switch back renders true colors); three direct units pin
exactly-once with extra render views attached.
2. Translation preserves the untouched fragments of a partially
overlapped span: left of the replaced range keeps its styling,
right of it shifts by the length delta, only the rewritten bytes
lose theirs (the writer styles what it writes; inserted bytes
inherit nothing). Pre-fix any overlap dropped the WHOLE span —
red abc, SGR reset, CR, X left bc unstyled; zero translation
painted the default X red instead. Bite: exact (glyph, fg) cells
X=default, b/c=red — any_styled_cell cannot see either failure.
3. The per-CR/BS/erase-line whole-prefix scan is gone:
slot.line_start is tracked — advanced at every \n (append helper
+ the mid-line newline branch), read O(1) by the rewind paths,
reset on run start/resync/raw marker appends. Measured on 2 MB of
output + 3000 CR updates (release): 2.52s pre-fix → 0.67s
post-fix (remainder is fixture-bound; pre-fix cost grows with
buffer size). No correctness bite is possible for a pure perf fix
— the committed test pins the tracked value's behavior across
multi-line appends, batch-boundary CR, repeated CR, erase-line,
and recovery paths, and passes on both implementations by design.
Gates: fmt; clippy workspace all-targets; lib 1531; crdt lib 1705;
compile acceptance 60; crdt acceptance 3; m4 101; m6.4 15; m6.5 11;
m6.8 8; GPU 59; workspace sweep 2517/0; git diff --check.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VoiEyuPjoBhvwACf8HAnLB
Levi Neuwirth
a49adc2589