:strict reference check: one pass over the bytes, no allocation

[mathieu17g]

Closes #142.

parse(…; wellformed = :strict) allocated about seven times per reference, in the Node and FlatNode parses alike. The check matched a regular expression with three groups per reference. Each match cost a RegexMatch and its captures, and a numeric reference a SubString for tryparse besides.

The check now reads the bytes of a token that carries a & and allocates nothing. Each & is found with findnext over the code units. A numeric reference has its digit run accumulated and tested against the XML Char range. A named one is compared byte for byte with the five predefined names, when the "Entity Declared" constraint applies. Only a rejected reference is copied, into its message.

What the check accepts, rejects and ignores is unchanged, message included. The digit run is read on the hexadecimal alphabet in both forms, as the pattern read it, so a decimal form carrying a letter is still rejected. A & that starts neither form is still left unchecked. The test suite pins this with a comparison against the pattern-based check, kept as the specification. It covers every string of one to four tokens over an alphabet that spells each reference form, each edge of the character range, a broken form and multibyte text: 406,900 strings. Each is checked as String and as SubString, names on and off, verdict and message compared. The edges are listed one by one as well. Allocation guards hold the check at zero, and parse(…; wellformed = :strict) at the allocation count of :structural on a document of references.

The decoder and the check now share one reference lexer. unescape resolves a reference through the same helpers, with its own outcomes, and the five names and the characters they denote sit in one place. Its outcomes are unchanged, pinned by the 579,194-string comparison of #144.

Measured

Section (8) of benchmarks/profile.jl, escaped twin of #140, 81,799 references; BenchmarkTools medians, Julia 1.12.7, final commit:

row	before	after
escaped :strict	89.8 ms (GC 13.5) · 3,241,570	68.9 ms · 2,658,494
escaped :structural (control)	56.8 ms · 2,658,494	56.4 ms · 2,658,494
plain :strict (control)	59.8 ms · 2,526,927	59.1 ms · 2,526,927
text-only :strict (control)	7.05 ms · 18	6.97 ms · 18

The acceptance line of #142 was escaped :strict at the allocation count of escaped :structural.

Per call, on a value of a spreadsheet cell's shape, A & B <tag>: 281 ns and 19 allocations → 17 ns and 0. On a 98-byte text token with one reference: 117 ns and 8 → 8 ns and 0. unescape on the same cell value: 33 → 34 ns, within the run-to-run spread. The shared-strings cells of benchmarks/benchmarks.jl move within noise: Node build 9.93 ms unchanged, normalized write 51.6 → 50.7, entity-heavy sst_load! 19.7 → 19.8.

Docs

• PERFORMANCE-v0.4.md: Table 7 republished from this campaign, its note dated per table; the sentence under it no longer describes a pattern match per reference. Table 6, benchmarks_results.md and the README table keep the figures of unescape: one pass over the bytes, one allocation per decoded value #144: nothing they measure depends on this change, and the same campaign re-measured them within noise.
• CHANGELOG.md: a Fixed entry.

Left out

_check_chars_strict, the character-range scan :strict runs on every token, is #143. entities.jl still resolves the character references of an entity's replacement text with its own regular expression, once per declaration, off the per-value path.

[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 96.61%. Comparing base (1448084) to head (f17f900).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #145      +/-   ##
==========================================
+ Coverage   96.57%   96.61%   +0.03%     
==========================================
  Files          15       15              
  Lines        2573     2597      +24     
==========================================
+ Hits         2485     2509      +24     
  Misses         88       88              

Files with missing lines	Coverage Δ
src/escape.jl	99.43% <100.00%> (+0.04%)	⬆️
src/parse.jl	99.40% <100.00%> (+0.04%)	⬆️

🚀 New features to boost your workflow:

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