xslt 1.6.5

This package has no dependencies.

Seven engine fixes, all found by running real stylesheets rather than the test suites. The
largest of them unblocked 60 of the 162 XSpec suites, which had been failing before they could
execute a single test.

### Fixes

- **`xsl:copy` carries the source element's complete IN-SCOPE namespaces, not only the `xmlns`
 declarations written on it.** XSLT 3.0 §11.9.1 requires a namespace node on the copy for every
 namespace node of the original. Emitting only the local declarations is usually
 indistinguishable, because the copy lands in a tree whose ancestors re-supply the rest — but
 not when nothing re-supplies them. Under `inherit-namespaces="no"`, or for a detached copy, the
 copy kept only the prefix of its own name, and a later `resolve-QName` against it failed with
 `FONS0004` for a prefix the source plainly had. XSpec's compiler is exactly this shape:
 `x:combine` wraps the combined document in `<xsl:element inherit-namespaces="no">` and the
 compiler then copies scenarios out of that tree and resolves `@function` / `@template` / `@as`
 against the copies. Measured across the XSpec suites: suites stuck at Compile 99 → 42, reaching
 Run 39 → 84, running to Completion 24 → 36, with 58 suites advancing a stage and none
 regressing.

- **`xsl:result-document` targeting the principal output returns the empty sequence.** Its content
 is a result document and never contributes to the containing sequence constructor (§26.1). One
 with an `href` was already redirected correctly; one targeting the principal output wrote into
 the same buffer an `as=`-typed body slices for its return value, so its content was counted as
 the template's result. A template declared `as="empty-sequence()"` around a bare
 `xsl:result-document` — the shape XSpec generates for its report — raised `XTTE0505: expected
 zero items, got 1`.

- **`xsl:result-document` content contributes no accumulated items either.** The companion to the
 above: the content ran with the enclosing sequence accumulator still installed, so an
 `xsl:sequence` *inside* a result document could still leak an item into an enclosing typed body.

- **The collected-attribute stack is restored bottom-first.** Four seams saved the stack to a
 `List`, cleared it, and pushed the list back — but `Stack<T>` enumerates top-first, so the
 restore reversed it. Harmless at nesting depth 1, which is why it survived; at depth ≥ 2 the
 enclosing element sealed against its parent's buffer, losing every attribute created by the
 `xsl:attribute` instruction while the parent silently gained them, and every element built
 afterwards was shifted by one. Literal and AVT attributes never enter that stack, so ordinary
 stylesheets were unaffected and only code generators — which must compute attribute names and
 values — fell in. It accounted for `XTSE0010` in 62 XSpec suites: a generated
 `<xsl:stylesheet>` with no `@version`, `Q{uri}` names with an empty local part, and valueless
 `xsl:attribute` elements were one defect wearing three masks.

- **`fn:transform`'s `initial-template` option is an `xs:QName`, not a local name.** It was read
 as a string, discarding the namespace, so a namespaced entry point could not be selected. The
 adjacent `initial-function` option already handled this correctly.

- **An `xsl:function` with no `as=` returns text NODES for text its body produces.** The typed
 form already did; the untyped form handed back a plain string, so `boolean()` over a sequence
 of such results raised `FORG0006`.

- **`for-each-group` with `group-by` over a grounded population is streamable.** The classifier
 reported every `group-by` as not-streamable, which forced whole-input buffering even where the
 population is already materialised. A correctness fix in the classifier; runtime behaviour is
 unchanged.

### Version alignment

From 1.6.5 the engine libraries and the CLI tools built on them share a single version, so a
matching number means a compatible set. See `PhoenixmlDb.Core` 1.6.5 for the rationale.