Move the anchor attribute in ANCH(n)/REC(n), not just its values; release 0.5.1

[shigarov]

Port of the jRegTab 0.5.1 fix (a092102, merged in 43c1fa9) — reference
implementation AnchorAttributeAtPosition.java.

The bug

apply_anchor_at_position (src/spec.rs) permuted only the record
values while the schema kept its original order. With anonymous $a_i names
(position == name) that is invisible, but with named attributes — the ones
AVP produces from a header row — the attribute-value binding broke: the column
carrying the anchor's name received another attribute's values.

Reported case: <ANCH(4)> over a Dato,Tid,Eksamen,Fagkode,Lokaler,Klasse
header came out as Lokaler,Dato,Tid,… with every value shifted.

The fix

The transformation now moves the anchor attribute — name together with its
values — so every record keeps its pairs and only the schema order changes,
making it the special case of apply_schema_reordering it always was.

One rule for named and anonymous attributes alike: an anonymous name is no
longer reassigned positionally, it travels with its attribute, so $a_1..$a_4
under ANCH(2) reads $a_2, $a_3, $a_1, $a_4 while the values stay in the
same positions as before. All ANCH/REC(n) task fixtures are header-less and
compare positionally — none was touched.

No change was needed in the RTL plumbing: <ANCH(n)>, inline REC(n) on an
atomic content specification and inline REC(n) inside a delimited one already
converge on the same Transformation. That is now pinned by a test rather than
left to inspection.

Tests

• #[cfg(test)] mod tests in src/spec.rs (the file had none): named,
  anonymous, mixed and degenerate schemas, a negative position, and a regression
  check that value positions are unchanged.
• tests/test_rtl_anchor_forms.py: the three forms must agree, and REC(n)
  under {','} yields one record per raw token (" C1.1" keeps its leading
  space — the 0.5.0 S_delim rule).
• AnchorAttributeAtPosition added to test_api.py's transformation coverage.
• Corpus re-pinned to jRegTab v0.5.1 (c126337) and synced byte-for-byte: the
  two new semantic cases anch_named_attrs and anch_named_inline_delim. Both
  set expectedHasHeader, which is what lets them see the bug at all.

cargo test --no-default-features    21 passed   (was 14)
pytest tests -q                   1925 passed   (was 1918)
tests/fixtures/                     untouched

Discriminating check: temporarily restoring the old return value fails 4 core
tests and 5 Python ones (both conformance cases included) while the regression
check stays green.

Docs and version

The ANCH(n) row in the settings table described an attribute-naming operation
that never existed; docs/rtl-reference.md now states what the transformation
does, including the anonymous-name rule and the fact that an inline REC(n) is
picked up anywhere in the pattern, delimited specifications included. Same
correction in the REC(n) operation row, atp.md, itm.md and api.md.

Version bumped to 0.5.1 across Cargo.toml, Cargo.lock, pyproject.toml,
__init__.py, README.md and docs/index.md. The README's differential note
still refers to the v0.5.0 run — that comparison was not repeated here.

Plan: plans/ANCH_MOVES_ATTRIBUTE.md (and a new plans/INDEX.md).

🤖 Generated with Claude Code