0.2.0

Correctness fixes to the pagination pipeline. All five were reproduced against a live probe harness before being fixed, and each carries a regression test verified to fail without its fix.

Fixed

• Short first page no longer inflates the page count (#9). The page count is derived from a new pageSize requirement on PagedResponse rather than from the first response's item count, so a page shortened by server-side filtering can no longer make the client request pages that don't exist — previously one out-of-range 404 discarded the entire load. The tail next_page is now adopted only from a page that actually contributed rows, so a clamped page can't trigger a full re-walk.
• De-duplication can no longer be silently disabled (#10). Conformers whose identity(of:) returns nil now take the sequential path, as the documentation always claimed. Previously they entered the parallel path with de-duplication off and could receive duplicate records.
• The parallel path is now bounded (#11). maxParallelPages (1000) mirrors maxSequentialPages, so a bogus server total can't turn one call into thousands of requests.
• A large total no longer crashes the process (#12). total is range-checked before the page-count arithmetic, surfacing a decode error instead of trapping on overflow.
• The pipeline genuinely runs off the main actor (#13). streamAllPages, perform, fetch, send, and fetchAllPages are now nonisolated. Previously the whole pipeline ran on the main actor despite the comments and Package.swift rationale claiming otherwise, hitching the main thread on large lists.

Upgrading

PagedResponse gains a pageSize static requirement with a default of 100. Existing conformers keep compiling, but a conformer whose real page size is smaller — or whose lists are shorter than 100 items — will fall back to the sequential walk. That is correct but slower. Declare your endpoint's actual page size to keep the parallel path.

Over-estimating pageSize is safe (the count comes out short and the remainder is walked); under-estimating is what the old code did and what #9 fixes.

Full changelog: 0.1.0...0.2.0