Add ignoreWhitespace search flag

[LazyCompiler]

Summary

Adds an opt-in ignoreWhitespace flag to SearchQuery. A literal query drops whitespace on both sides of the fold instead of collapsing it, so invoice finds the letter-spaced i n v o i c e that OCR'd scans and tracked-out headings produce, and total amount finds totalamount.

Live demo: https://lazycompiler.github.io/embedpdf-ignore-whitespace-demo/ — a static build of this branch (@embedpdf/engine, wasm in a worker) against the fixture that ships with the change. Search invoice, then tick ignoreWhitespace, then wholeWord; try total amount.

Design

• A fold option, not a new matcher. FoldOptions.dropWhitespace skips whitespace units in foldText; the existing map already carries hits back to exact original ranges, so a match spans the page's whitespace. The default fold is unchanged (no SEARCH_FOLD_VERSION bump). Like matchCase / matchDiacritics the flag is a non-default fold: SearchReader re-folds the raw page text per query and canonicalSearchQuery passes the query through.
• wholeWord boundaries on the original text. With whitespace dropped, i n v o i c e 42 folds to invoice42, so a boundary check on the folded plane would reject a genuine whole word. Under ignoreWhitespace, isWholeWordHit reads the boundaries off haystack.original around the mapped range; the default path is untouched.
• Literal-only. regex: true + ignoreWhitespace is rejected by validateSearchQuery with a new issue ignore-whitespace-with-regex, mirroring matchDiacritics (regex runs on the raw text plane; a pattern spells its own \s*).
• Identity / wire. Included in the cursor query key (searchCursor.ts, CloudDocumentSearchService.queryIdentity), the zod SearchQuerySchema, the search token (SearchTokenSchema + encode/decode), and the cloud search route (ignoreWhitespace=true).
• Not exposed in the viewer-chrome search panel here (same stance as matchDiacritics); happy to add a toggle if wanted.

Tests

• Unit (engine-core): fold (drop + decomposition-born whitespace), literal (spaced page text, spaced needle, glued words, line wraps, matchCase and wholeWord interplay, whitespace-only needle), regex validation, canonicalSearchQuery, token round-trip.
• Conformance (runSearchConformance): the flag can only add hits over the default fold and the space-free needle finds the same hits; the regex combo is rejected with InvalidArg.
• Real engine: packages/engine/main/test/search-ignore-whitespace.test.ts runs against test/fixtures/letter_spaced_text.pdf — a 2 KB fixture authored by a committed deterministic generator (fixtures/generate-letter-spaced-fixture.mjs, documented in fixtures/README.md, following the action_payloads.pdf pattern). It covers letter-spaced hits on both pages with drawable geometry, a word split across a line break, a spaced needle finding a glued word, wholeWord boundaries from the original text, and matchCase composition. Happy to drop the fixture and keep only the unit + conformance tests if you'd rather not grow the fixture set.
• Verified locally against the wasm runtime: the full @embedpdf/engine suite (45 files / 332 tests) and engine-core (436 tests) pass; engine-services and @cloudpdf/engine typecheck.

Changeset included (minor: @embedpdf/engine-core, @embedpdf/engine-services, @cloudpdf/engine, @cloudpdf/server).

[vercel]

@LazyCompiler is attempting to deploy a commit to the CloudPDF Team on Vercel.

A member of the Team first needs to authorize it.