[AmitMY]

Simplifies pixel_renderer/renderer.py (−79/+40 lines) and makes short-text rendering ~12% faster, with pixel-identical output.

Changes

• One growable canvas per block_size. This replaces the fixed 1024px reusable surface plus a fresh surface/context for every wider text. Long texts now reuse their surface too, and the two-branch logic is gone.
• Fewer cairo calls per render. The black source is set once when a canvas is created. The white background is filled through the numpy view of the surface (mark_dirty() before drawing, flush() before reading).
• @cache'd layout getter in place of the global _measurement_context tuple.
• Dropped a redundant np.ascontiguousarray in render_signwriting and fixed the docstring's block_size default (16, not 32).

Benchmarks

Ubuntu 24.04 container matching CI (pango 1.52.1, cairo 1.18.0, Python 3.12), aarch64, best of several runs:

	main	this PR
benchmark.py word mix	~110k it/s	~123k it/s (+12%)
long_text_benchmark.py (512 words)	~1,125 it/s	~1,141 it/s (+1.5%)

Correctness

• 4,420 renders compared against main's render_text: regression samples, benchmark words, the long lorem text, empty / newline / 3,000-char / SignWriting inputs, at sizes (16,12), (32,24), (16,12), (24,16) interleaved so canvases get both reused and grown. 0 mismatches, same exceptions.
• Full suite passes (458), including the pixel-exact backward-compatibility tests.

Where the rest of the time goes (why this isn't faster)

Profiled per call:

• Long text: Pango layout/shaping ~53%, drawing ~23%, BGRX→RGB ~18%.
• Short words: after this PR, ~1.4µs of the ~7.9µs is conversion and ~4µs is Pango.

I tried and rejected:

• Faster BGRX→RGB conversion. byteswap + V3 views, uint32 shifts and PIL's BGRX unpacker were all slower than the existing three strided assignments, which run close to memory bandwidth.
• Rendering to an A8 surface. It breaks color emoji (Noto Color Emoji is in the font set).
• Hand-driving Pango.itemize/shape. Fragile through GI, and unlikely to stay pixel-identical (bidi reordering, etc.).

🤖 Generated with Claude Code

https://claude.ai/code/session_01Fy1wuWdqcKuebgj3St8dME

---

Note

Low Risk
Performance refactor in rendering only; PR reports pixel-identical output across thousands of regression renders and full tests.

Overview
Refactors Pango/Cairo text rendering in renderer.py for speed and simpler reuse: one @cache’d shared layout replaces the old measurement-context globals, and one growable canvas per line height (with a cached numpy pixel view) replaces the fixed 1024px surface plus a separate path for wide text.

render_text now clears the needed width via pixels.fill(255) and mark_dirty(), draws once, then flush() before bgra_to_rgb; long strings reuse the same surface after it grows. render_signwriting drops redundant ascontiguousarray, and the block_size doc default is corrected to 16.

^{Reviewed by Cursor Bugbot for commit c582f77. Bugbot is set up for automated code reviews on this repo. Configure here.}