bug(common/models): Error in ModelCompositor.predict causes keyboard failure

[keyman-sentry]

Sentry Issue: KEYMAN-ANDROID-GS

Fatal keyboard error with obolo_chwerty:obolo_chwerty for  language. Loading default keyboard.

The breadcrumb shows:

packageID: obolo_chwerty
keyboardID: obolo_chwerty
keyboardType: SYSTEM
cm_message: Uncaught TypeError: Cannot read property 'sample' of undefined
cm_sourceID: blob:null/2eaf2e9d-75fc-490e-9426-446345506678
keyboardName: Obolo Chwerty
cm_lineNumber: 170

This was a little tricky to track down as we don't yet have good source visibility into the predictive text worker. I used the Chrome Web Inspector to view the blob in minified form, and there was only one reference to sample on line 170, which resolved to ModelCompositor.predict.

So, we appear to be getting an error in ModelCompositor.predict - perhaps an empty array or malformed data?

keyman/common/predictive-text/worker/model-compositor.ts

Lines 78 to 80 in 2b66926

	let inputTransform = transformDistribution.sort(function(a, b) {
	return b.p - a.p;
	})[0].sample;

Note that this error is occurring with various keyboards, including sil_euro_latin, and not just obolo_chwerty. I have linked related crash reports where I can find them.

[keyman-sentry]

Sentry issue: KEYMAN-WEB-11

[keyman-sentry]

Sentry issue: KEYMAN-ANDROID-11F

[keyman-sentry]

Sentry issue: KEYMAN-ANDROID-14Z

[keyman-sentry]

Sentry issue: KEYMAN-ANDROID-GS

[keyman-sentry]

Sentry issue: KEYMAN-ANDROID-14V

[keyman-sentry]

Sentry issue: KEYMAN-ANDROID-128

[keyman-sentry]

Sentry issue: KEYMAN-ANDROID-142

[keyman-sentry]

Sentry issue: KEYMAN-ANDROID-GS

[jahorton]

Closest thing I can currently do for a repro: I built a temporary unit test for @keymanapp/input-processor to replicate the condition that could give rise to the error: an empty array of Alternates. The test case's code:

        let languageProcessor = new LanguageProcessor();
        languageProcessor.init();

        let contextSource = new com.keyman.text.Mock("li", 2);
        let transcription = contextSource.buildTranscriptionFrom(contextSource, null, null);
        // This should break things.
        transcription.alternates = [];

        languageProcessor.loadModel(modelSpec).then(function() {
          languageProcessor.predict(transcription).then(function(suggestions) {
            assert.isOk(suggestions);
            assert.equal(suggestions[0].displayAs, '«li»');
            assert.equal(suggestions[0].transform.insert, ' ');
            assert.equal(suggestions[1].displayAs, 'like');
            assert.equal(suggestions[1].transform.insert, 'like ');
            done();
          }).catch(done);
        }).catch(function() {
          assert.fail("Unexpected model load failure");
          done();
        });

The result:

Commenting out the line that "should break things" indeed unbreaks this temporary test case.

So, the remaining question is... what's causing an empty alternates array? That said... this could allow us to add 'breadcrumbs' of a sort outside of the worker, since it can be detected within KMW proper.