For each element in a plane, predicts its value from the *previous* values using a small fixed-context-model hash table, then emits `truth XOR prediction`. The output is concentrated around zero (heavy leading-zero distribution) when the underlying values change smoothly — exactly the regime where Huffman / [`crate::codecs::Fpc`] both win. ## Predictor: fixed-context model (fcm) Maintains a hash table of `1 << HASH_BITS` entries indexed by a rolling hash of recent values. On each step: ```text prediction[i] = table[hash] residual[i] = value[i] XOR prediction[i] table[hash] = value[i] hash = ((hash << SHIFT) ^ value[i]) & ((1 << HASH_BITS) - 1) ``` The first element has an all-zero hash context, so the prediction is the table's initial (zero) entry and the residual is the value itself. From then on the predictor learns local patterns. The decoder runs the same recurrence in lock-step to recover the original. ## Single-predictor variant Burtscher's original FPC uses *two* predictors (fcm + dfcm) and emits one bit per element selecting the better. PTWM ships only the fcm half here — the dual-predictor variant adds a per-element side channel that would need its own plane and complicates the chain shape; the single-predictor version captures most of the gain on trained-weight exponent streams and stays byte-exact-invertible through a single byte plane. ## Element widths Same support matrix as [`crate::transforms::IntDelta`]: `Byte`, `Word2`, `Word4`, `Word8`. `Nibble` is rejected.