fix: replace force-unwraps with guards in TextDecoder.decodeText

[yw5aj]

Fixes #414

Two force-unwraps in decodeText crash the app when decoder state degrades during sustained rapid transcription calls (e.g., live transcription preview at ~5 calls/sec):

• Line 747: decoderInputs.initialPrompt.last! — crashes when initialPrompt is empty
• Line 829: decoderOutput.logits! — crashes when logits is nil

This PR replaces them with guard statements that throw descriptive errors, allowing callers to handle the failure gracefully instead of crashing the entire app.

Minimal diff — no behavioral change for normal operation

The guard statements only trigger when the values are unexpectedly nil. Normal transcription flow is unaffected.

[ZachNagengast]

Interesting, third PR this week for this specific issue #420 #417. I see you are the author of the original issue, do you have any relation to the other two authors?

[yw5aj]

No relation — we're an independent team that hit the same crash during live transcription. Glad to see others confirming the issue. Happy to defer to whichever PR you prefer to merge!

[ZachNagengast]

👍󠄳󠅜󠅟󠅣󠅕

[achyutbenz19]

Took an independent look at the diff. A couple of things that might be useful.

1. Rebase is clean. The branch is on 2870b46, and main has moved 6 commits ahead, most notably #456 (the package rename) and #455 (the swift-transformers to in-tree Hub/Tokenizers switch). git rebase main applies cleanly with no conflicts. Post-rebase, the patch is a pure +8/-2 on Sources/WhisperKit/Core/TextDecoder.swift. Confirmed with Swift 6.2 / Xcode 26 on Apple Silicon: the rebased HEAD builds, and the model-free TextDecoder tests (testTextDecoderMLMultiArrayInputType, testTextDecoderMLMultiArrayOutputType, testTextDecoderMLMultiArrayOutputTypeWithNilValues, testDecodingCacheInitialization, testDecodingCacheWithNilValues, testMLMultiArrayConformsToTextDecoderTensorType) pass 6/6 on both main and the rebased commit. Commands:

git fetch origin
git checkout fix/safe-unwrap-text-decoder
git rebase origin/main

2. Adjacent pattern, noting for your judgment. Lines 843 and 844 of current main (sampleResult.tokens.last!, sampleResult.logProbs.last!) are the same shape as the two unwraps this PR fixes. #420 covers them, this PR does not. Not suggesting you expand scope, just flagging in case it changes the merge calculus.

Disclosure: I am an AI assistant (Anthropic's Claude) helping a user contribute to WhisperKit. I did not reproduce the crash deterministically, it depends on rare runtime states that a file fixture will not reliably hit, so this is build and test signal only, not before / after empirical evidence.