XPath 2.0 phase 4: the numeric type hierarchy — instance of recognizes xs:integer, xs:decimal and xs:float as well as xs:double, tracked for literals and explicit casts
Differential and generated testing against the ISO reference implementation
Fuzz targets, criterion benchmarks, clippy pedantic, corpus test suite, runnable examples, and the specification
Five more lints — unused variables, empty rules, empty patterns, duplicate assertion tests, phases that activate nothing — and rule shadowing generalised from three special cases to pairwise subsumption
extends href, with fragment identifiers on both it and include
A denial of service in nested ranges and for loops, found by fuzzing, closed with a budget shared across every nested construct in one expression
Three optimisations found by profiling: linear location building, rule claims in a vector rather than a hash map, and a fused walk for the common rule context
XPath 2.0 phases 5 through 9: the remaining sequence functions, cardinality assertions and atomization, deep-equal(), resolve-uri(), and the adjust-*-to-timezone() family
XPath 3.0 phase 1: function items — inline function expressions, named function references, dynamic calls — and a real for-each(), the one function they exist for
XPath 3.0 phase 2: the arrow operator => and string concatenation ||
A stack overflow in eight repeating binary operators, found by fuzzing, closed with a shared parse_binary_chain helper
XPath 3.0 phase 3: the rest of the higher-order sequence functions — filter, fold-left, fold-right, for-each-pair — and function-item introspection
XPath 3.0 phase 4: the simple map operator !, plus the let expression found while documenting it
EQNames, Q{uri}local, for node name tests — the other gap phase 4 found, plus a stack-overflow regression in the recursion-depth margin, caught by fuzzing and closed the same day
Union types retracted as a gap, not implemented as one — neither the XPath 3.0 nor 3.1 grammar has a (T | T) union-type literal at all, so the syntax this roadmap once cited was never real XPath to begin with
Streaming validation, requested by the maintainer, overriding an earlier deferral here — one repeating record parsed, matched, fired, and discarded at a time, for schemas provably local to it, with no rework of the arena or NodeId model after all: verified with a counting allocator that live bytes stay identical at 10,000, 100,000, and 1,000,000 records
One item remains, and it is an argument for not doing the
work yet. That is deliberate: a roadmap that only lists ambitions is a wish
list.
1. no_std core — blocked, and the earlier reasoning was wrong
The claim was that only I/O and the resolver need std. They
are not the obstacle: quick-xml, which this crate's XML
parser is built on, declares no no_std support and reaches
for std::io and std::error throughout. Nothing
in the tree can be parsed without it, so a no_std build would
validate nothing.
regex, the dependency that looked likelier to block it, turns
out to be #![no_std] already and would only need its std feature dropped.
So the real price is a hand-written XML tokenizer replacing a
heavily-fuzzed one — a correctness risk taken in a validator,
whose whole value is being right — bought for a target that WebAssembly,
the plausible use case, already reaches with std. Not worth
it on today's evidence.
Examined and abandoned
Recorded so they are not proposed again
A lint for a context that can never match
There is no vocabulary to check against. Schematron declares no element
names of its own; it is layered on a grammar — a DTD, XML Schema, RELAX
NG — that it cannot see. Inferring one from the names the schema happens
to mention is circular: a context naming an element no test mentions is
entirely normal. This would need the grammar as a second input, which is
a different feature.
A lint for a report that reads like a requirement
Confusing assert with report is the classic
Schematron mistake, and catching it would be valuable. But the only
available signal is the wording of English prose, and a lint that
misfires on "this invoice must be reviewed manually" teaches its reader
to ignore the linter — the one outcome worth avoiding above all.
Not planned
Out of scope, on purpose
A roadmap item here is not a refusal to discuss it — it is a record of the
reasoning, so a new argument can be weighed against the old one instead of
restarting it.
