stet 0.5.0 — 64-bit integers, Type 3 font fixes, --page

Minor release. PostScript integers are now 64-bit, which fixes the standard
LCG idiom that programs use for pseudo-randomness and is the reason this is a
breaking release rather than a patch. Three Type 3 font defects and a
clippath coordinate-space bug are also fixed, and the CLI gains --page.

Breaking

Downstream Rust code that reads PostScript integers needs attention; nothing
in the PostScript language surface changed incompatibly.

• PsValue::Int now carries i64 instead of i32. A
  match obj.value { PsValue::Int(v) => … } binds an i64, so any use site
  that needs an i32 no longer compiles. DictKey::Int and Token::Int
  widened with it, as did Context::rand_seed.
• PsObject::as_i32() now range-checks. It returns None for a value
  outside i32, where before it always returned Some for an integer. This
  is the one change with no compiler error behind it — audit call sites that
  treat None as "not an integer". Use the new as_i64() for the full
  range; keep as_i32() where the value is genuinely bounded (array and
  string indices, character codes), since a too-large value should fail those
  callers' range checks rather than wrap into a valid-looking index.
• PsObject::int() takes impl Into<i64>. Calls are unaffected; only
  code coercing it to a fn(i32) -> PsObject pointer breaks.

Fixed

• Type 3 fonts supplying only BuildGlyph raised invalidfont. The show
  path required BuildChar unconditionally and pushed the character code.
  PLRM 5.7 lists BuildGlyph as preferred and makes BuildChar required only
  "for LanguageLevel 1 or if BuildGlyph is absent", so such a font is
  well-formed and must be handed the character name from Encoding.
  Ghostscript renders these; stet refused them. xshow/yshow/xyshow had
  the identical defect.
• stringwidth raised invalidfont on every Type 3 font, BuildChar
  ones included. It branched for font types 2, 0 and 42 and then fell through
  to the Type 1 path, which looks for CharStrings — a Type 3 font has none.
  There is no width table to consult: the width is whatever the build
  procedure hands setcachedevice/setcharwidth, so the procedure now runs
  inside a gsave/grestore with its marks drained, and measuring paints
  nothing.
• glyphshow raised invalidfont on every Type 3 font. It read
  FontType but never branched on 3, going straight to the CharStrings
  lookup. Per the PLRM it now invokes BuildGlyph with the name directly —
  bypassing Encoding, which is what lets glyphshow reach glyphs no
  character code maps to — or, with only BuildChar, reverse-searches
  Encoding for the name and pushes the array index, retrying with
  /.notdef and raising invalidfont only when neither is encoded.
• PostScript integers are now 64-bit, matching Ghostscript, which fixes the
  standard LCG idiom PostScript programs use for pseudo-randomness:
  /seed seed 1103515245 mul 12345 add 2147483648 mod def. On 32-bit integers
  the product overflowed, promoted to a real, and mod — which is
  integer-only — raised typecheck. Widening only the real fallback would not
  have fixed it: the product needs 55 bits and a real carries 53, so the seed
  would have come out one too high and every later draw would have diverged
  silently from what other interpreters produce. PLRM Appendix B's 32-bit
  range is listed under "Typical Limits" for interpreters "running on 32-bit
  machines" which "do not necessarily apply to all PostScript
  implementations", so this is not a conformance change. Overflow past the
  64-bit range still promotes to a real. bitshift is correspondingly 64-bit
  wide, and cvi accepts the wider range.
• clippath returned the page in device space instead of current user
  space, so a program that had transformed its coordinate system got a clip
  rectangle dragged along with the transform. The clippath fill idiom for
  painting a background then filled an offset region and left part of the page
  bare — visible in the tiger EPS, whose grey backdrop was displaced by its
  %%BoundingBox origin. The default clip is a fixed region of the device, so
  it is now derived with the default CTM; pathbbox and fill map it back
  through the current CTM, which is what puts it in user space for the caller.
  Ghostscript's values now match exactly under translate, scale and rotate.
• cvi on a long integer-valued string was off by one. The string scanner
  returned f64, so (22358003463039195) cvi came back as ...196 after the
  round trip through a 53-bit mantissa. Integer literals now stay integral.

Added

• The CLI reports a page that was painted but never shown. A program that
  paints marks and then ends without a matching showpage leaves them on a
  page the device is never asked to emit, and the page is discarded. That is
  correct — it is what the PLRM specifies and what Ghostscript's file devices
  do — but it was indistinguishable from a broken renderer: no file appeared
  and nothing said why. The warning distinguishes a program that produced no
  output at all from one that lost only its trailing page.
• Interpreter::warnings() surfaces the same diagnostic to library
  callers, where the silence was worse: render() returned Ok(vec![]), an
  empty page list that reads as a legitimate result. New public types
  ExecWarning and ExecWarningKind in stet::diagnostics; the CLI shares
  the detector, so the two cannot drift. Programs that install nulldevice
  are exempt — that is the PLRM-sanctioned way to ask for no output, so marks
  left unemitted are the point rather than a mistake.
• --page sets the page size for PostScript/EPS input — a named size
  (letter, legal, tabloid, ledger, executive, a0-a6, b4, b5)
  or WIDTHxHEIGHT in points, with an optional -landscape / -portrait
  suffix that swaps the dimensions. There was previously no way to render a
  plain %!PS program whose artwork is larger than the default page:
  %%BoundingBox sets the page only for EPS — for a non-EPS document DSC
  makes it a description of the artwork's extent, not a page-size request, so
  both stet and Ghostscript fall back to US Letter and clip. --page
  overrides an EPS %%BoundingBox when both apply, and is rejected for PDF
  input, whose pages carry their own size.
• GlyphCache::by_type3_name, a name-keyed Type 3 glyph cache. glyphshow
  can name a glyph that no character code maps to, which leaves nothing for
  the existing code-keyed cache to key on.

WebAssembly

• stet-wasm 0.2.0. Its JavaScript API is unchanged, but the browser
  build inherits everything above, so rendering output moves: clippath
  backgrounds fill the page, Type 3 fonts that previously raised
  invalidfont render, and PostScript programs using the standard LCG for
  pseudo-randomness run instead of failing. A minor rather than a patch
  because the pixels change, not because anything you call does.