XPath 1.0 / XSLT 1.0 spec-derived coverage matrix#

Date: 2026-07-17 · Trigger: owner escalation, 2026-07-17 —

"I can't show this work proudly to ANYONE if you hide massive holes like that. Why didn't it show in w3c tests?"

The namespace:: axis was unimplemented while summaries reported XSLT "near-perfect". Root cause: coverage was measured against self-selected test corpora and self-authored unit tests, not against the specifications. The XSLT corpus is a subset of w3c/xslt30-test filtered to only the features slice-1 already implements (see §Corpus), and the XPath unit suite contains a test that asserts the hole is a passing feature (tests/unit/xpath_tests.ml:273 — check_bool "namespace axis is deferred: clean parse failure" true (parse_fails "namespace::*")). A corpus can never surface a hole it was built to exclude; a unit test that pins the gap green hides it.

This document enumerates the specifications — every XPath 1.0 axis, every §4 core function, every XSLT 1.0 element/attribute/function — and marks each: implemented? (F* file:line), exercised by any test?

🔴 Holes summary (read this first)#

Emoji register: 🔴 unimplemented + user-visible · 🟡 unimplemented corner · ⚠️ implemented-but-partial / untested-for-correctness · ✅ implemented + exercised.

~25 spec features unimplemented; +5 partial/stubbed. The corpus and unit suites exercise essentially none of the holes.

# Feature Spec Sev Why it matters
1 namespace:: axis XPath §2.2 (axis 13/13) 🔴 The escalated hole. namespace::node() selects nothing; parser rejects the axis outright.
2 translate() XPath §4.2 core string fn 🔴 Common case-folding / char-map idiom. Returns "" (falls through to unknown-fn default).
3 lang() XPath §4.3 core boolean fn 🔴 xml:lang language-range test. Unimplemented → "".
4 xsl:number XSLT §7.7 🔴 List/section numbering. Emits nothing.
5 xsl:key + key() XSLT §12.2 🔴 Cross-reference lookups. Neither element nor fn exists.
6 format-number() + xsl:decimal-format XSLT §12.3 🔴 Numeric formatting. Returns "".
7 xsl:import / xsl:include / xsl:apply-imports XSLT §2.6 🔴 Stylesheet composition + import precedence. Top-level collector ignores them.
8 xsl:processing-instruction XSLT §7.3 🔴 PI output. Falls to else -> [].
9 xsl:message XSLT §13 🟡 Diagnostics / terminate.
10 generate-id() XSLT §12.4 🔴 Node-identity keys. Missing.
11 system-property() XSLT §12.4 🔴 xsl:version/xsl:vendor probes. Missing.
12 unparsed-entity-uri() XSLT §12.4 🟡 DTD unparsed-entity lookup. Missing.
13 disable-output-escaping (on value-of/text) XSLT §16.4 🔴 Raw markup injection. Ignored.
14 use-attribute-sets + xsl:attribute-set XSLT §7.1.4 🟡 Shared attribute bundles. Ignored.
15 xsl:strip-space / xsl:preserve-space XSLT §3.4 🟡 Source whitespace control. Ignored.
16 xsl:namespace-alias XSLT §7.1.1 🟡 Aliasing (e.g. generating XSLT). Ignored.
17 xsl:fallback + forwards-compatible processing (version) XSLT §2.5, §15 🟡 Graceful degradation of unknown instructions. Ignored.
18 xsl:output attributes beyond method XSLT §16 🟡 indent, encoding, doctype-*, cdata-section-elements, omit-xml-declaration. Only method read.
— element-available() XSLT §15 ⚠️ Hard-coded false for all names, incl. supported instructions.
— document() XSLT §12.1 ⚠️ Only the empty-string (document("") = stylesheet) form; any real URI → empty node-set.
— current() XSLT §12.4 ⚠️ Correct at expression top level; inside a predicate returns the step context, not the XSLT current node (self-disclosed divergence, XPath.Eval.fst:1394).
— function-available() XSLT §15 ⚠️ Knows XPath core + document/current only; reports false for key/format-number/generate-id/system-property (consistent with their absence, but not spec-complete).

Test-visibility of the holes: 0 of the 30 rows above is exercised by the 88-test XSLT corpus or the XPath unit suite. namespace:: is "tested" only by an assertion that it fails to parse.

1. XPath 1.0 axes (§2.2) — 12 of 13 implemented#

Parser: Parser.XPath.axis_of_name (Parser.XPath.fst:268-281). Evaluator: XPath.Eval.apply_axis (XPath.Eval.fst:718-731).

Axis Parsed? Evaluated (file:line) Test coverage Status
child ✅ child_axis :438 xpath-unit, XSLT corpus ✅
descendant ✅ descendant_axis :443 xpath-unit descendant:: ✅
parent ✅ parent_axis :422 xpath-unit parent:: ✅
ancestor ✅ ancestor_axis :430 xpath-unit (4×) ✅
following-sibling ✅ :466 xpath-unit (4×) ✅
preceding-sibling ✅ :471 xpath-unit (2×) ✅
following ✅ following_axis :703 xpath-unit (1×) ✅
preceding ✅ preceding_axis :711 xpath-unit (1×) ✅
attribute ✅ attribute_axis :448 xpath-unit, XSLT avt/axes ✅
self ✅ :720 (apply_axis) xpath-unit (5×) ✅
descendant-or-self (//) ✅ :723 xpath-unit, XSLT match ✅
ancestor-or-self ✅ :726 xpath-unit ✅
namespace ❌ axis_of_name→None :281 no Ax_Namespace constructor asserted-to-fail only (xpath_tests.ml:273) 🔴

Root cause note (from Parser.XPath.fst:281): "namespace: deferred (Parser.XML has no namespace-node kind)". Parser.XML's xml_node models no namespace-node, so the axis has no node kind to yield. This is a data-model gap, not a one-line parser addition. The evaluator does carry full namespace-URI machinery (resolve_ns_uri, elem_ns_uri, namespace-uri()) for name tests — but the axis that materialises namespace nodes as selectable/countable items is absent. node-1601 (namespace::node() → attribute list) is the one W3C test that hits it; it is the sole XSLT-corpus fail.

2. XPath 1.0 core function library (§4) — 25 of 27 implemented#

Dispatch: XPath.Eval.eval_funcall (XPath.Eval.fst:1331-1486).

§4.1 Node-set (7/7): last :1336, position :1335, count :1337, id :1344, local-name :1370, namespace-uri :1381, name :1370 — all ✅ (xpath-unit + XSLT id/function-available categories; id() DTD-ID landed 2026-07-17).

§4.2 String (9/10): string :1403, concat :1407, starts-with :1415, contains :1408, substring-before :1420, substring-after :1427, substring :1438, string-length :1450, normalize-space :1453 — ✅. translate — 🔴 absent (no branch; hits else XV_Str "" :1486). Not in is_supported_xpath_function :1091.

§4.3 Boolean (4/5): boolean :1460, not :1456, true :1458, false :1459 — ✅. lang — 🔴 absent.

§4.4 Number (5/5): number :1462, sum :1466, floor :1470, ceiling :1472, round :1474 — ✅ (xpath-unit math).

Grammar-production families (all parsed, Parser.XPath.fst): LocationPath (abs/rel, // desugar), Step (axis::/abbrev/@/./..), NodeTest (name/*/ns:*/text()/comment()/node()/processing-instruction()), Predicate, PrimaryExpr ($var/(expr)/literal/number/FunctionCall), FilterExpr, all binary/unary operators, Union |. PI-target node test processing-instruction('t') parses (NT_PI (Some t)) and evaluates (:631), but XSLT.Transform drops PI alternatives from a union select before eval (self-disclosed, XSLT.Transform.fst:86) — ⚠️.

3. XSLT 1.0 elements (§ per row) — 22 of ~35 implemented#

Instruction dispatch: XSLT.Transform.fst:1444-1545 (ln = "…" chain, else -> []). Top-level collectors: templates :1673, globals collect_globals :1716, output method :1693.

Element Impl (file:line) Test Status
xsl:stylesheet/xsl:transform :1786 corpus ✅
xsl:template (match/name/mode/priority) :1673-1680 corpus ✅
xsl:apply-templates (select/mode/sort) :1466 corpus ✅
xsl:call-template :1483 — (blocklisted in corpus selection) ✅ (untested by corpus)
xsl:with-param bind_with_params :1492 — ✅ (untested)
xsl:param / xsl:variable (local+global) :1381, :1716 corpus variable ✅
xsl:value-of :1446 corpus ✅
xsl:for-each (+sort) :1456 corpus select ✅
xsl:sort (data-type/order, stable) :1209 corpus sort (24) ✅
xsl:if :1450 corpus ✅
xsl:choose/when/otherwise :1454, :1597-1602 corpus ✅
xsl:element (name/namespace AVT) :1509 corpus construct-node ✅
xsl:attribute :1537 corpus avt ✅
xsl:text :1448 corpus ✅
xsl:comment :1541 corpus node ✅
xsl:copy instantiate_copy :1507 corpus copy ✅
xsl:copy-of (+copy-namespaces) :1495 corpus copy ✅
xsl:output :1693 (only method) corpus ⚠️ partial
xsl:number ❌ (else->[]) — 🔴
xsl:key ❌ — 🔴
xsl:import/xsl:include ❌ (not collected) — 🔴
xsl:apply-imports ❌ — 🔴
xsl:processing-instruction ❌ (else->[]) — 🔴
xsl:message ❌ — 🟡
xsl:decimal-format ❌ — 🟡
xsl:namespace-alias ❌ — 🟡
xsl:attribute-set ❌ — 🟡
xsl:strip-space/xsl:preserve-space ❌ — 🟡
xsl:fallback ❌ — 🟡

4. XSLT 1.0 attributes of significance#

Attribute Impl (file:line) Status
match, name, select, test throughout ✅
mode (template + apply-templates) :1467, :1679 ✅
priority tpl_prio :1680 ✅
namespace (xsl:element) :1520 ✅
copy-namespaces (2.0) :1499 ✅
data-type/order (xsl:sort) :1214 ✅
disable-output-escaping ❌ (comment-only :81) 🔴
use-attribute-sets ❌ 🟡
version / forwards-compatible processing ❌ 🟡
xml:space / whitespace stripping ❌ 🟡
case-order/lang (xsl:sort collations) ❌ (disclosed :75) 🟡

5. XSLT 1.0 defined functions (§12) — 2 of ~9 complete#

Function Impl (file:line) Status
document() :1357 — empty-string form only ⚠️ partial
current() :1394 — top-level only ⚠️ partial
function-available() :1476 — XPath-core + document/current ⚠️ partial
element-available() :1480 — hard-coded false ⚠️ stub
key() ❌ 🔴
format-number() ❌ 🔴
generate-id() ❌ 🔴
system-property() ❌ 🔴
unparsed-entity-uri() ❌ 🟡

Corpus provenance audit#

Where the 88-test suite came from (third_party/testing/xslt/README.md): a vendored subset of w3c/xslt30-test @ fddf1cf920087e791f13315d68dfbe874d97dc56 (cloned 2026-07-07), W3C-test-suite licence. 64 tests across 18 categories +24 xsl:sort tests (2026-07-08) = 88.

The selection criteria are the root cause. README §"Selection criteria" #4 vendors a test only if every xsl:* element it uses is in the slice-1 allowlist, with an explicit blocklist: import/include, call-template, sort, number, key, function, apply-imports, next-match, …, message, …, decimal-format, namespace-alias, attribute-set, fallback and no mode attribute. Any test that would exercise a §Holes row was filtered out at vendoring time. The suite is structurally incapable of surfacing the holes — it is a regression pin for what already works, not a conformance measurement. (The sort category was later added when xsl:sort landed, i.e. the corpus grows to match the engine, not the spec. That is backwards for hole-detection.)

The full available corpus. Two established XSLT-1.0 conformance collections exist beyond xslt30-test:

  1. OASIS XSLT/XPath conformance test suite (the historical reference set, ~1900–2400 cases, contributed by Lotus/IBM, Microsoft, Oracle/Sun, and others). It is the suite node-1601 itself originates from (the <?spec xpath#axes?> PI and the node-NNNN/match-NNNN/copy-NNNN naming are OASIS conventions — xslt30-test re-vendored many OASIS cases). Licensing: OASIS test materials carry the OASIS IPR / a per-contributor licence; the Xalan-hosted copy is Apache-2.0. This suite has dense coverage of exactly our holes: axes (namespace axis), number, key, impincl (import/include), attribset, namespace-alias, output attributes, string (translate), boolean (lang).
  2. Apache Xalan tests/ conformance set (Apache-2.0, git-clonable) — a cleaned, runnable mirror of the OASIS cases plus Xalan-specific additions. Most tractable to vendor offline (no HTML-only endpoints).

Recommendation (concrete). Import a hole-targeted slice — selection driven by the §Holes matrix, not by the current allowlist — from the Apache-2.0 Xalan mirror (avoids the OASIS IPR question):

Until then, the honest headline is "XSLT: 87 pass, 1 fail on a slice-1-scoped subset of xslt30-test; ~25 spec features unimplemented, untested by that subset" — never "near-perfect".

Dispositions-vocabulary rule (added to the ledger)#

New rule (also added to docs/claude-rules/w3c-completeness-ledger.md §Dispositions):

disputed-fixture and by-design may only be applied when the underlying capability EXISTS. A missing feature is always planned-family/gap, stated as such. Labelling a hole "implementation-defined" or "disputed fixture" implies the processor made a correct choice among valid behaviours — which is false when the processor produces nothing because the feature was never built.

Re-audit against the new rule#

Row Current label Verdict
XSLT node-1601 (ledger line 88) "dispositioned (implementation-defined namespace-node order / unmodeled namespace:: axis)" 🔴 MISLABEL. "implementation-defined namespace-node order" implies namespace nodes are produced, just ordered differently. Reality: namespace:: is unmodeled → zero nodes, output <NSlist/>. Per the new rule this is planned-family(#302, branch xslt-namespace-nodes)/gap. The XSLT detail section (line 600) already correctly calls the family "🟡 in flight / planned" — the summary row contradicts its own detail. Fix the summary row's wording.
XSLT score row 88 vs detail 593 Row: "87 pass, 1 fail". Detail header: "79 pass, 9 fail" ⚠️ Internal inconsistency (row updated 2026-07-17, detail section stale at 79/9). Not a disposition mislabel; flagged for the score-sync pass (out of scope for this docs-only commit — no score change).
GRDDL "6 fail-known-gap-xslt-feature", "30 fail-graph-mismatch" planned-family(#301) ✅ Correct — missing XSLT-fidelity features labelled as gap, not by-design.
rml_io 7+4+1+1 skips dependency-blocked(named component) ✅ Correct — each names an absent engine (XPath-in-RML, decompression, RDBMS, SPARQL iterator). Missing capability, honestly labelled.
rml_io RMLSTC0009a planned-family(no issue) ✅ Rule-compliant (gap, flagged as needing triage; capability question is error-detection, which the runner has in shape).
eecc 47 by-design(non-VCDM serializations) by-design ⚠️ Defensible but soft: SD-JWT/mdoc/JWT-VC are different formats, out of the VCDM-JSON-LD suite target — closer to "out-of-suite data" than "feature we lack". A no-parser-for-format case borders on gap; recommend re-phrasing to by-design(out-of-suite serialization; not a VCDM-JSON-LD credential) to make the "not our target" claim explicit rather than implying we chose not to verify a credential we could.
eecc 4 crypto skips dependency-blocked(live did:web + no key material) ✅ Correct.
xml_conformance 386+332+416+37 skips by-design(Stage-A scope limits) ✅ Defensible — a non-validating XML processor is a conformance class the XML 1.0 spec explicitly defines (§5.1), so declining DTD-validation tests is a correct-by-design behaviour, not a hidden hole. The runner's "parsed-but-not-validated" wording is transparent about it. Rule-compliant because the label does not claim the missing validation was implemented. (Watch item: if a summary ever rounds this to "XML: green", that would become a mislabel — the 1171 skips must stay visible. Cross-ref docs/designissues/2026-07-17-xml-conformance-skip-census.md.)
OWL profile EL/QL 🧭 untriaged ✅ Honestly marked untriaged, not falsely dispositioned.
OWL dl-909, FS2RDF-literals-ar, I5.5-005 disputed-fixture ✅ Genuine fixture disputes (the checker capability exists; the vendored graph/label is contested). Rule-compliant — these are the legitimate use of disputed-fixture.

Net: one clear mislabel to correct (node-1601 summary wording), one score-consistency flag (XSLT row vs detail), one soft re-phrasing suggestion (eecc). No other gap-class mislabelings found — the other suites correctly reserve by-design/disputed-fixture for scope-classes and genuine fixture disputes where the capability exists.