feat(xsd): transcode extract_classes.py to Rust — byte-faithful, closes XSD↔TTL bijection, drops the Python oracle dep

The MARS XSD classification extractor (arago/MARS-Schema/tools/extract_classes.py,
~360 lines: ~140 extraction logic + ~150 table formatting) is now a
faithful Rust transcode at crates/ogar-from-schema/src/xsd.rs, behind
an optional `xsd` feature (pulls roxmltree, pure-Rust read-only XML
DOM; the default TTL path stays zero-parser-deps).

Not huge — the transcode is ~350 LOC Rust including tests. And it
doubles as the seed of the broader XSD → Class front-end: the same
walk that extracts classifications is the structural-arm lift for any
XSD schema.

What it lands:

* BYTE-FOR-BYTE transcode proof. xsd::to_asciidoc() reproduces the
  Python `-F asciidoc` output exactly — 628 lines, including verbatim
  XSD-documentation whitespace and the printAsciiDocFooter trailing
  newline. Test xsd::tests::asciidoc_matches_python_oracle diffs
  against the cached _oracle/classifications.adoc.

* XSD↔TTL BIJECTION CLOSED (was "queued" in MARS-TRANSCODING.md §2).
  xsd::tests::xsd_classes_match_ttl_enum asserts full bidirectional
  set-equality between the XSD-extracted Application value set and the
  TTL validation-parameter enum — not just one-directional membership.
  Two independent encodings of one taxonomy, provably equal both ways.

* PYTHON DEPENDENCY REMOVED from the calibration path. `cargo test
  --features xsd` is the whole oracle now; no python3 interpreter
  needed. extract_classes.py stays vendored in _oracle/ as the
  provenance witness (what the transcode was proven against), not a
  runtime dep.

Transcode discipline (faithful to the Python semantics):

* getAttribute("xml:lang") returns "" for absent (not None); lang
  filter is "absent OR en". roxmltree resolves xml: to the xml
  namespace, matched on attribute.name() == "lang".
* getXMLText concatenates DIRECT text-node children only (not
  recursive); the documentation's internal whitespace is load-bearing
  for the byte-match.
* :revdate: is datetime.now() in Python (non-deterministic); the Rust
  to_asciidoc(c, revdate) takes it as a parameter so output is
  reproducible and testable.
* The two-level extension chain (master complexType carries NodeType,
  intermediate complexType carries Class, leaf element carries
  SubClass) + the post-process phase that stitches base→element is
  reproduced exactly, including the master_types gate.

Tests: 20/20 with --features xsd (16 default + 4 new xsd); 16/16 on
default (xsd code fully feature-gated). Clippy-clean (--no-deps),
fmt-clean.

Docs:

* docs/MARS-TRANSCODING.md §2 — bijection marked closed; Python-dep
  removal noted; the Rust extractor added to the oracle-direction table.
* .claude/board/EPIPHANIES.md — FINDING with the transcode-discipline
  notes for the next source→Rust port.

Author: claude

.claude/board/EPIPHANIES.md

@@ -15,6 +15,58 @@
 
 ## Entries (newest first)
 
+## 2026-06-22 — extract_classes.py transcoded to Rust byte-faithfully; XSD↔TTL bijection closed; Python dependency removed from the oracle
+**Status:** FINDING
+**Scope:** XSD front-end × calibration self-containment × the queued bijection
+
+The MARS XSD classification extractor (`arago/MARS-Schema/tools/extract_classes.py`,
+~360 lines, ~140 logic + ~150 table formatting) is now a faithful Rust
+transcode at `crates/ogar-from-schema/src/xsd.rs`, behind the optional
+`xsd` feature (pulls `roxmltree`, a pure-Rust read-only XML DOM; the
+default TTL path stays zero-parser-deps).
+
+Three things this lands:
+
+1. **Byte-for-byte transcode proof.** `xsd::to_asciidoc()` reproduces
+   the Python `-F asciidoc` output exactly — 628 lines, including the
+   verbatim XSD-documentation whitespace and the `printAsciiDocFooter`
+   trailing newline. Test: `xsd::tests::asciidoc_matches_python_oracle`
+   diffs against the cached `_oracle/classifications.adoc`.
+
+2. **The XSD↔TTL bijection is closed (was "queued" in
+   `MARS-TRANSCODING.md §2`).** `xsd::tests::xsd_classes_match_ttl_enum`
+   asserts FULL bidirectional set-equality between the XSD-extracted
+   Application value set and the TTL `validation-parameter` enum — not
+   just one-directional membership. The XSD and the TTL are two
+   independent encodings of one taxonomy and they now provably agree
+   in both directions.
+
+3. **The Python dependency is removed from the calibration path.**
+   `cargo test --features xsd` is the whole oracle now; no `python3`
+   interpreter needed. `extract_classes.py` stays vendored in
+   `_oracle/` as the provenance witness (what the transcode was proven
+   against), not a runtime dep.
+
+Transcode discipline notes (for the next source→Rust port):
+- The Python `getAttribute("xml:lang")` returns `""` for absent (not
+  `None`); the lang-filter is "absent OR en". roxmltree resolves `xml:`
+  to the xml namespace — match on `attribute.name() == "lang"`.
+- `getXMLText` concatenates DIRECT text-node children only (not
+  recursive); the documentation's internal whitespace is load-bearing
+  for the byte-match.
+- The `:revdate:` is `datetime.now()` in Python (non-deterministic);
+  the Rust `to_asciidoc(c, revdate)` takes it as a parameter so the
+  output is reproducible and testable.
+
+Answer to "is it huge": no — ~360 lines, half output formatting; the
+transcode is ~350 LOC Rust including tests. And it doubles as the seed
+of the broader XSD→`Class` front-end (the same walk that extracts
+classifications is the structural-arm lift for any XSD).
+
+Evidence: `crates/ogar-from-schema/src/xsd.rs` (20/20 tests pass with
+`--features xsd`; 16/16 on default). `docs/MARS-TRANSCODING.md §2`
+updated to mark the bijection closed.
+
 ## 2026-06-22 — OGIT is the canonical template store; Odoo (and any source-AST producer) digests INTO it; consumers relive agnostically via askama
 **Status:** FRAMING
 **Scope:** Foundry-parity collapse × cross-consumer architecture × digest-once-relive-N

crates/ogar-from-schema/Cargo.toml

@@ -11,14 +11,21 @@ description = "Schema-as-input producer family for OGAR IR. Lifts the STRUCTURAL
 [features]
 default = []
 serde = ["dep:serde", "ogar-vocab/serde"]
+# The `xsd` feature pulls in a read-only XML parser for the XSD
+# front-end (the faithful Rust transcode of arago/MARS-Schema's
+# extract_classes.py). Kept optional so the default TTL path stays
+# zero-parser-deps.
+xsd = ["dep:roxmltree"]
 
 [dependencies]
 ogar-vocab = { path = "../ogar-vocab" }
 serde = { workspace = true, optional = true }
+roxmltree = { version = "0.20", optional = true }
 
-# No external TTL/XSD parser pulled in yet: the v0 reader uses a
-# narrow line-oriented walker that handles the shapes OGIT's TTL
-# actually emits (a tiny, machine-stable subset of full Turtle).
-# When the surface grows (Wikidata-shaped TTL, full RDF/XML, OWL
-# imports), swap in `oxttl` / `oxrdf` here without touching the
+# The default TTL reader uses a narrow line-oriented walker that
+# handles the shapes OGIT's TTL actually emits (a tiny, machine-stable
+# subset of full Turtle) — no external parser. The `xsd` feature adds
+# roxmltree (pure-Rust, read-only DOM) ONLY for the XSD front-end.
+# When the TTL surface grows (Wikidata-shaped TTL, full RDF/XML, OWL
+# imports), swap in `oxttl` / `oxrdf` without touching the
 # Class/Attribute target shape this crate produces.

crates/ogar-from-schema/src/lib.rs

@@ -60,6 +60,8 @@ use ogar_vocab::{Attribute, Class, EnumDecl, EnumSource, Language};
 pub mod sgo;
 pub mod ttl;
 pub mod ttl_emit;
+#[cfg(feature = "xsd")]
+pub mod xsd;
 
 /// What a single TTL file describes — exactly one of: an entity (`Class`),
 /// a datatype attribute (`Attribute`), or a verb (`Association` shape).