[Velli20]

Summary

Removes the Operands type from pdf-content-stream-operators. Operators now decode their operands with the existing pdf-object-reader decoders, reading in place from the parser's reused operand buffer.

• PdfOperator::read takes &mut OperandContext (an ArrayContext over a ReadSession) and returns ReadResult.
  • Matrices are Self::new(operands.read_array()?), numbers operands.at(i)?, and names/strings operands.at::<PdfString>(i)?.bytes, which shares the parsed bytes.
  • TJ uses the Vec<PdfTextItem> decoder and sc/SC use read_all.
• The parser keeps one Vec<ObjectVariant> per stream and passes it as ArrayContext::new(&operands, session). No per-operator Vec, Arc or byte copies.

pdf-object-reader changes that make this faster than before

• ReadSession::read decodes direct values in place instead of copying each one into a new PdfObject (Arc) and cloning the traversal state. Following references moved into read_reference.
• ObjectContext borrows &ObjectVariant, so context.object().value() becomes context.object() across the decoder crates.
• ObjectVariant::kind() and ObjectVariant::dictionary() now hold the logic; PdfObject and ResolvedObject delegate to them.
• ArrayContext borrows any slice (new ArrayContext::new) and gains read_array::<T, N>(). read_all presizes its Vec.
• Adds #[inline] on the small hot decoders. The release profile (opt-level = "z", fat LTO), which benches inherit, barely inlines without it.

Benchmark

crates/pdf-content-stream/benches/operator_parsing.rs, release profile. Each of 5 interleaved rounds ran a git archive export of main and this branch back to back.

group	main	this branch	change	faster in
dispatch_heavy	493.7 µs	461.8 µs	−6.5%	5/5
mixed	2316 µs	2235 µs	−3.5%	5/5
numeric_heavy	1302 µs	1255 µs	−3.6%	5/5

Test plan

• cargo fmt --check
• cargo clippy --all --workspace
• cargo test --workspace --exclude pdf-graphics-skia --exclude pdf-conformance: 1,110 passed. Skia crates not built locally; CI covers them.

🤖 Generated with Claude Code