A bare except handler hid thousands of unreadable tables behind an 86.1 percent recall score

2,868 of 7,892 is 36.3 percent ^[8]. The post puts the missing share at roughly 39 percent ^[6]. The larger count fits better: 3,056 of 7,892 is 38.7 percent, and 3,056 is what the post reports for a 183-character checkout root, against 2,868 for a 182-character one ^[7][8]. One extra character in a directory name cost another 188 table files ^[9]. The loader walked the checkout with `path.rglob("*.json")` and wrapped each read in `except Exception: continue` ^[2]. `rglob` returns the path. `open` refuses it, because Windows will not open a path over 260 characters ^[3]. What comes back is `FileNotFoundError` for a file that is sitting on disk, and the handler counted every one of those tables as a table that does not exist ^[5]. "A crash is loud. A harness that prints 86.1% when it measured the wrong thing is not, and I published every one of those numbers," the author wrote ^[23]. The held-out evaluation is the more instructive failure. Six fixtures get built. One of them declares three extra schemas with four objects in them: three tables all named `account`, plus a view joining across two, and that collision is most of what the fixture exists to test ^[16]. The harness opened a bare SQLite connection with no `ATTACH`, so `executescript` raised `unknown database "billing"` on the first of those statements. The run died there, after the five per-schema rows had printed and before the TUNE, HELD OUT and OVERALL lines ^[17]. According to the post, that file had never printed an aggregate at all ^[18]. `ATTACH DATABASE ':memory:'` before the DDL clears the exception: the script completes, the harness prints its aggregate, exit code 0 ^[19]. In-memory databases are dropped when the connection closes, and the harness closes that connection before `bootstrap` opens a new one ^[20]. The aggregate then reports 256 objects indexed, all in `main`, and no objects named `account` ^[20]. "The repair converted a loud failure into a quiet one, and both of us verified the repair by reading a number that the repair had not changed," the author wrote ^[21]. The arrangement that works is file-backed databases attached on a connect listener, with the schema list passed to bootstrap ^[22]. One check reported success while being incapable of failing. The reviewer reported that 52 of 52 questions had a step where the returned-set size decreases as K grows, and the author's check reported 0 of 52 ^[13]. That check compared `returned(K)` against `returned(K-1)`, and since top-K picks are nested and both expansion passes are unions, `returned` cannot decrease in K, so the test could never fire ^[14]. Comparing the overages instead of the totals gave 52 of 52, matching the reviewer ^[15]. One of the six defects is documentation. `Scored.reason` documented five values, `hybrid | vector | lexical | fk | pinned`, and the selector writes a sixth, `covers`, so a caller branching on the field falls through silently on one case in six ^[12]. Forty lines above the block that does the displacing sits the comment "Bounded, and only ever additive -- it cannot displace a ranked pick". The code had been displacing a ranked pick since a commit made a day and a half earlier ^[10][11]. The comment shipped in three releases ^[11]. This is one developer's account of his own harness, published on dev.to, and nobody in the record has reproduced the counts independently ^[25]. For the 2,868 figure to say anything about another benchmark, three conditions have to hold: the run executes on Windows, where `open` refuses paths over 260 characters ^[3]; the loader enumerates the corpus and swallows read errors instead of counting them ^[2]; and the corpus has paths deep enough to cross the limit from wherever it was cloned ^[7]. Spider 2.0's checkout meets the third at a 182-character root ^[7].