OffsetIncrement.next_page_token re-extracts the full page to count records, discarding the retriever's last_page_size

[bazarnov]

Symptom

Every declarative stream paginated with DefaultPaginator + OffsetIncrement parses and extracts each response page twice: once in the retriever (record emission) and a second time inside the pagination strategy, purely to count records.

Root cause

At airbyte-cdk 7.3.1, airbyte_cdk/sources/declarative/requesters/paginators/strategies/offset_increment.py (next_page_token, lines ~78-92):

decoded_response = next(self.decoder.decode(response))

if self.extractor:
    page_size_from_response = len(list(self.extractor.extract_records(response=response)))
    last_page_size = (
        page_size_from_response if page_size_from_response is not None else last_page_size
    )

• self.extractor.extract_records(response=response) re-runs the full record extraction and materializes the whole page into a list just to take len().
• The wiring is unconditional for declarative sources: model_to_component_factory.py passes extractor_model=model.record_selector.extractor into create_default_paginator >> create_offset_increment, so the paginator always holds a second instance of the stream's extractor.
• The incoming last_page_size argument - already counted per emitted record by SimpleRetriever - is effectively always overwritten: len() never returns None, so the fallback branch is dead.

Impact

Double CPU and a transient second in-memory copy of every page on the pagination hot path. For large-page APIs this is material: source-google-analytics-data-api pages are up to 100,000 rows / ~45 MB JSON per page, and the same connector's extraction path was measured around 1.17 GiB RSS per 100k-row page during incident triage - the counting pass adds a second full materialization on top.

Reproduction

Any manifest with DefaultPaginator + OffsetIncrement; add a counter or log line to the record extractor - it fires twice per page (once from the retriever, once from next_page_token).

Suggestion

Count from what the retriever already extracted (it computes last_page_size per record) instead of re-parsing the raw response, or make the extractor-based counting opt-in for the nested-record cases it was added for.

Precedent

Found while reviewing airbytehq/airbyte#83188 and airbytehq/airbyte#83318 (GA4 pagination fix: CursorPagination >> OffsetIncrement on 25k-100k row pages).