What did you get wrong that took longest to find?

[akash-coded]

War stories, specifically the ones where the system was working perfectly and the results were
wrong. Those are the expensive ones, because nothing errors and no alert fires.

Three from this repository, all real, all in the tracker:

	What happened	How long it hid
#1	The FTS5 tokenizer split ERR_CONN_RESET into three terms present in every document. Query matched everything, ranked nothing. Zero errors.	Only caught by a by-class slice; the aggregate barely moved
#2	A pure k-NN graph with no long-range links. recall@20 = 0.00 at ef=64 — not degraded, structurally impossible	Looked like an embedding problem
#7	A subsample taken for speed moved the null base rate from 15% to 33%, and abstention precision is base-rate sensitive. The chart contradicted the prose above it	The prose was right the whole time

What to include

The detection gap is the valuable part. Not "what broke" but "why did we find out the way we
did, and what measurement would have caught it sooner?" That is the last line of every
postmortem in the incident runbook,
and it is the one people skip.