:strict character-range scan: bytes, not characters

[mathieu17g]

Closes #143.

The :strict character-range scan decoded every character of every text, attribute value, comment, CDATA section and processing-instruction body, about 1 ns per byte. It now reads the bytes. An ASCII byte is legal or not by itself, so only a byte at or above 0x80 is decoded, by the string's own Char iteration. The verdict, the message and the error on malformed UTF-8 are those of the loop it replaces.

Two paths, by span length. A span of 16 bytes or less, an attribute value or the white space between two tags, is checked as two 8-byte words without a loop. A longer span is checked 64 bytes at a time, in a loop the compiler vectorises.

Section (8) of benchmarks/profile.jl on Julia 1.13.0, medians, GC share in parentheses:

parse(…, Node; wellformed = …)	before	after
text-only :structural	0.49 ms	0.50 ms
text-only :strict	7.54 ms	0.81 ms
plain :structural	48.44 ms	49.07 ms
plain :strict	62.98 ms (GC 7.5)	54.32 ms (GC 6.9)
escaped :structural	57.74 ms (GC 3.4)	58.0 ms (GC 3.3)
escaped :strict	78.38 ms (GC 13.4)	64.22 ms (GC 7.8)

Allocations are unchanged: :strict allocates what :structural does, and a guard in test/test_allocations.jl holds the scan to zero on both paths, for a String, a SubString and a StringView.

The issue asked for plain :strict within 2 ms of plain :structural; measured: 5.3 ms between the medians, all of it within the 6.9 ms GC share of the :strict run, so net of GC the two do not separate. What remains is a cost per span, not per byte: the document has 623,553 spans, half of them white space of a few bytes between tags, and each one costs the call and the two-word test.

Tests: the loop the scan replaces is recopied as the specification and compared with the scan on every code point, on 18,420 byte sequences of an alphabet covering lead, continuation, control and boundary bytes, and on 28,908 placed cases. Each case runs on a String, on a SubString through both paths and on a StringView.

Docs: Table 7 and the well-formedness paragraphs of PERFORMANCE-v0.4.md re-measured on Julia 1.13.0; CHANGELOG entry.

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 96.67%. Comparing base (b9c005c) to head (dd0b05b).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #148      +/-   ##
==========================================
+ Coverage   96.61%   96.67%   +0.06%     
==========================================
  Files          15       15              
  Lines        2597     2649      +52     
==========================================
+ Hits         2509     2561      +52     
  Misses         88       88              

Files with missing lines	Coverage Δ
src/parse.jl	99.54% <100.00%> (+0.13%)	⬆️

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.