Reading a JSON array as a string held every Facebook Page to three posts

The cursor was in the response the whole time, two layers from where the code reached for it. The page-info chunk is identified by its Relay label, a string ending `..._timeline_list_feed_units$page_info`, and the cursor sits at `data["page_info"]["end_cursor"]`, not `data["end_cursor"]`, according to the write-up ^[6]. Two wrong assumptions were stacked, the type of `path` and the depth of the cursor, and both had to be found before pagination worked at all ^[6].

A failed `isinstance` test does not raise. `extract_next_cursor` read `line["path"]`, tested `isinstance(path, str)`, and returned `(None, False)` when the test failed ^[4]. That return is also what an exhausted feed looks like, so the walker stopped after one refetch page per Page and ignored `max_posts_per_page` ^[5].

The unit suite could not catch it. The NDJSON files that exercised the parser were hand-written with `path` as a dotted string and the cursor fields flat under `data`, which is the shape the buggy parser expected ^[8]. The suite was checking the parser against input written to the parser's expectations. Those files have been replaced with two live captures from facebook.com/nasa, trimmed of irrelevant bookkeeping and not reshaped ^[9].

The end-to-end test missed it because it never asked for more than the ceiling. An earlier run requested 2 Pages at 10 posts each, hit the same 3-per-page limit, and returned 6 rows, which satisfied a lenient "SUCCEEDED with some rows" check ^[7]. The post calls that 15% of the request; 6 rows against 20 requested is 30% ^[1]. The deliberate deep run of 3 Pages at 40 posts each returned 3, 3 and 3 ^[1].

Billing is per event: $0.20 per run start plus $0.003 per post landed, quoted as $3.20 per 1,000 posts ^[12]. That quote assumes one run start amortised across 1,000 posts. Under the ceiling, a run across three Pages lands 9 posts and bills $0.227, about $25 per 1,000 posts ^[2]. After the fix the same run lands 120 posts and bills $0.56, about $4.67 per 1,000 ^[3]. The measured platform cost of that 120-row run was roughly $0.0356, about $0.30 per 1,000 rows ^[13].

The same guessed-shape error hit four post fields. `text`, `reaction_count`, `share_count` and `comment_count` were read from top-level node keys the real payload does not have, and the engagement counts actually sit several `comet_sections` layers down ^[10]. One field was checked and left alone: `is_cross_post` read False on every post in the sample, the actor-id-versus-page-id comparison behind it is correct as written, and none of the sampled NASA or Nike posts were cross-posts ^[11].

The verification covers three Pages at one moment, with NASA, Nike and Microsoft returning 40, 40 and 40 after the fix, for 120 distinct post IDs ^[3]. What the new code keys on is a label string and a field path, so the result transfers to other Pages only if those Pages return the same refetch response shape. No account or access token is involved; the scraper reads the public timeline anyone can see logged out ^[15].