New defaults for -Zlr/-Zmc; possibly 5% gain in compression.

this was a huge oversight because both -Zlr and -Zmc configures
the adaptation and yet we never tried to tweak -Zmc until now!
JS code seems very non-stationary, thus smaller -Zmc fares better.

cli.js

@@ -65,9 +65,9 @@ Output options:
 -Zco|--context-bits BITS [Range: 1..24+, Default: derived]
   Sets the size of each context model, as opposed to the total size (-M).
   The maximum can range from 24 to 30 depending on the number of contexts.
--Zlr|--learning-rate RATE [Range: 1..2^53, Default: 256]
+-Zlr|--learning-rate RATE [Range: 1..2^53, Default: 500]
   Configures the learning rate of context mixer; smaller adapts faster.
--Zmc|--model-max-count COUNT [Range: 1..32767, Default: 63]
+-Zmc|--model-max-count COUNT [Range: 1..32767, Default: 5]
   Configures the adaptation speed of context models.
   Context models adapt fastest when the context is first seen,
   but become slower each subsequent occurrence of the context.

index.js

@@ -536,10 +536,10 @@ export class Packer {
             sparseSelectors: options.sparseSelectors ? options.sparseSelectors.slice() : defaultSparseSelectors(),
             maxMemoryMB: options.maxMemoryMB || 150,
             precision: options.precision || 16,
-            modelMaxCount: options.modelMaxCount || 63,
+            modelMaxCount: options.modelMaxCount || 5,
             recipLearningRate:
                 options.recipLearningRate ||
-                Math.max(1, Math.round((options.learningRateDenom || 256) / (options.learningRateNum || 1))),
+                Math.max(1, Math.round((options.learningRateDenom || 500) / (options.learningRateNum || 1))),
             contextBits: options.contextBits,
             arrayBufferPool: options.arrayBufferPool,
             numAbbreviations: typeof options.numAbbreviations === 'number' ? options.numAbbreviations : 64,