Keep the line breaker's per-atom fast path free of mixed-style atom handling

[nicoburns]

LLM Contributions: Generated with Opus 5.5 High

Fixes a linebreaking performance regression from #864 ("Handle line heights changing within atoms").
My LLMs break_all_lines+align benchmark on a ~250KB uniform-style English document went 3.27ms → 3.98ms (+22%).

DRAFT because benchmarking is currently showing that #855 gains all the benefits of this change and more

Root Causes

• BreakerState::append_atom_to_line stopped inlining into break_next and became an out-of-line call per atom
• LineBoxMetrics::add_text did extra work per atom before the last_text early-return: an eager atom.characters() slice (with bounds check) and a data.runs[run_idx].has_mixed_style_atoms load.

Fixes

• Pass the atom's characters as atom.char_range() and only slice ShapedText::characters() on the slow path.
• Encode the has_mixed_style_atoms into (item_idx, style_index). add_text_boxes sets it to the NO_LAST_TEXT sentinel after adding an atom of a run with mixed-style atoms, so such runs never hit the early return. The fast path check becomes just last_text == (item_idx, style_index) again.

Benchmarked "break" phase:

• 8af7c03 3.27–3.33ms,
• main 3.98–4.04ms
• This PR 3.26–3.31ms

Changelog: None

[tomcur]

My LLMs break_all_lines+align benchmark on a ~250KB uniform-style English document went 3.27ms → 3.98ms (+22%).

Interesting! For me 5794c746 (#864) vs 8af7c030 (#863, that PR's parent, which was the performance side of this) still benches close to neutral.

Line Breaking - arabic 8000 characters, narrow     [ 206.7 us ... 205.7 us ]      -0.49%
Line Breaking - arabic 8000 characters, wider      [ 107.1 us ... 108.9 us ]      +1.64%*
Line Breaking - arabic 8000 characters, unwrapped  [  78.4 us ...  80.8 us ]      +3.08%*
Line Breaking - latin 8000 characters, narrow      [ 172.8 us ... 172.0 us ]      -0.46%
Line Breaking - latin 8000 characters, wider       [  82.4 us ...  83.3 us ]      +1.04%*
Line Breaking - latin 8000 characters, unwrapped   [  52.2 us ...  53.5 us ]      +2.58%*
Line Breaking - japanese 8000 characters, narrow   [ 215.1 us ... 214.7 us ]      -0.23%
Line Breaking - japanese 8000 characters, wider    [ 143.2 us ... 145.5 us ]      +1.58%*
Line Breaking - japanese 8000 characters, unwrapped [ 121.0 us ... 121.6 us ]      +0.53%
Line Breaking - arabic 8000 characters, wrapped + spacing [ 116.3 us ... 118.7 us ]      +2.01%*
Line Breaking - latin 8000 characters, wrapped + spacing [  92.3 us ...  93.2 us ]      +0.98%
Line Breaking - japanese 8000 characters, wrapped + spacing [ 148.6 us ... 150.0 us ]      +0.98%
Page Break - arabic 8000 characters                [  86.2 us ...  88.9 us ]      +3.17%*
Page Break - latin 8000 characters                 [  61.1 us ...  62.9 us ]      +3.01%*
Page Break - japanese 8000 characters              [ 133.6 us ... 133.2 us ]      -0.31%

And versus that PR's parent's parent 97beea38, quite positive:

Line Breaking - arabic 8000 characters, narrow     [ 230.1 us ... 210.9 us ]      -8.31%*
Line Breaking - arabic 8000 characters, wider      [ 123.9 us ... 112.1 us ]      -9.57%*
Line Breaking - arabic 8000 characters, unwrapped  [  92.7 us ...  82.4 us ]     -11.10%*
Line Breaking - latin 8000 characters, narrow      [ 183.5 us ... 176.1 us ]      -4.05%*
Line Breaking - latin 8000 characters, wider       [  87.3 us ...  84.4 us ]      -3.27%*
Line Breaking - latin 8000 characters, unwrapped   [  55.5 us ...  54.4 us ]      -1.99%*
Line Breaking - japanese 8000 characters, narrow   [ 234.6 us ... 218.6 us ]      -6.82%*
Line Breaking - japanese 8000 characters, wider    [ 154.2 us ... 145.2 us ]      -5.86%*
Line Breaking - japanese 8000 characters, unwrapped [ 129.8 us ... 122.7 us ]      -5.44%*
Line Breaking - arabic 8000 characters, wrapped + spacing [ 133.1 us ... 119.2 us ]     -10.46%*
Line Breaking - latin 8000 characters, wrapped + spacing [  95.4 us ...  93.0 us ]      -2.53%*
Line Breaking - japanese 8000 characters, wrapped + spacing [ 160.6 us ... 150.3 us ]      -6.44%*
Page Break - arabic 8000 characters                [  99.0 us ...  88.2 us ]     -10.88%*
Page Break - latin 8000 characters                 [  64.2 us ...  62.8 us ]      -2.05%*
Page Break - japanese 8000 characters              [ 143.6 us ... 135.0 us ]      -6.00%*

[tomcur]

#855 that landed gives a further nice speed-up on our benches. What exactly is your break_all_lines+align doing?

[nicoburns]

My benchmarks are showing something like -8% from #863, then +18% from #864, then -25% from #855 (which for some reason seems to subsume the benefits of this PR). I think it's partly due to inlining decisions. So it may depend on compiler settings (notably codegen-units) and the phase of the moon....

[nicoburns]

(not necessarily intending to land this)