Flat Recall Is a Lottery at Ten Million Vectors — A Recall-vs-Bytes Map of Vector Compression, Read Under Two Rulers

A recall-vs-bytes map of vector compression from 0.01 to 200 bytes per vector across four datasets, read under two rulers: flat recall (the field's standard) and engine semantics (shortlist by code, re-rank exactly — the semantics of IVF, DiskANN, and ScaNN). At ten million items the flat ruler misreads codecs (a near-perfect codec scores 0.13 flat recall@10; 0.95 under the engine ruler). The map has structure: saturation widths that move with pool size (16B at 1M, 75B at 10M), a naming bound at log2(N) bytes below which codes filter rather than describe, and a tiny-code zone at 10M scale where a plain per-dimension codec beat Google's published RSLM reference in 10 of 11 byte-matched settings under the flat ruler (tie at display precision, no losses; tie-scale parity under the engine ruler). The map then designed its own engines: a twin family whose recall reproduces the map to within 3e-4 across five platforms from a watch to a cloud VM. At 100M vectors: 12.5 bytes retrieves 0.076, 25 bytes retrieves 0.905. Ten map cells were pre-registered under hash-witnessed commitment: seven landed in their locked intervals, two soft falsifications, one hard. Companion to the Yellow Phoenix trilogy (edge, cascade, watch, Mac-class papers).