Simplify stream boundary recovery

[Velli20]

Summary

• Collapse the three stream recovery entry points into one recover_stream_range(length: Option<usize>). parse_stream_recovering and parse_stream_recovering_direct_length become thin wrappers that differ only in how /Length is read. skip_stream_recovering is removed; the linear object scanner calls the core method directly, so it no longer copies the stream.
• Replace the code that moved the parser and restored it on failure with pure functions over the input, built around a small StreamTerminator { start, end } type.
• Speed: with a stale /Length hint, the terminator scan stops at the first valid candidate past the declared boundary instead of scanning to the end of the file, since every later candidate lies farther away. Previously, recovering a document whose streams had wrong lengths took time quadratic in the file size.
• Simplify line-ending trimming to a suffix strip, drop a dead special case for a match at the very start of the data, and match the truncated endstrea with one optional m.

endstream stays required. endobj marks where the object ends, but only endstream marks where the data ends, and requiring both is the main guard against matches inside binary stream data.

Apart from the speed fix, parsing gives the same results. stream_recovery.rs goes from 334 to 233 lines.

Test plan

• cargo test -p pdf-parser -p pdf-document
• cargo clippy: no new warnings compared with the base commit
• Read-only conformance on the PDFium corpus: 840 documents, 0 read worse, 0 read better than the base commit
• Read-only check of the pdf.js corpus (left to CI; needs a local PDFium library)

🤖 Generated with Claude Code