fix(adaptive): guard None history in convergence improvement delta

[wilke0818]

Problem

apply_convergence_marks (adaptive/convergence.py) computed the per-bucket round-over-round improvement as hist[-2]["aggregated_uncertainty"] - hist[-1]["aggregated_uncertainty"] with no None guard.

aggregated_uncertainty is float | None — a bucket is None when the axis can't score that round. The utterance axis marks a bucket comparison_status="incomparable" (→ None) when a window has no comparable ASR words, common on sparse/gappy speech. A bucket can flip None↔float across adaptive rounds (reprocessing adds/removes words), so its history can hold a None next to a float, and the subtraction raises TypeError: unsupported operand type(s) for -: 'NoneType' and 'float' — crashing the loop for the whole clip, even though the current value is already guarded a few lines up (if u is None: continue).

Impact: adaptive_loop failed on 5/30 clips in a real batch — all sparse-speech (Word-color-Stroop, Random-Item-Generation, Story-recall, Picture-description), which have many incomparable utterance buckets.

Fix

Guard both history entries; if either round was None, improvement is unmeasurable across it, so leave improvement=None (bucket stays open, not stalled) instead of crashing. No behavior change when both rounds are scored.

Testing gap

Same blind spot as the recent subprocess-venv fixes: the adaptive tests don't exercise a clip whose cross-round history contains incomparable (None) buckets. A fixture with a None→float history transition would regression-cover it.

Base: alpha.

🤖 Generated with Claude Code

Version

Published prerelease version: 1.3.1-alpha.37

Changelog

🐛 Bug Fix

• fix(adaptive): guard None history in convergence improvement delta #541 (@wilke0818)
• refactor(cache): cached-inference layer out of analyze_audio.py (T051, parts 1-2) #538 (@satra)
• Scene-aware presence/quality axis + ASR transcription overhaul (CrisperWhisper 2.0, Qwen) #536 (@satra @claude)
• feat(canary): split long audio at pauses to avoid the ~40s truncation #534 (@wilke0818)
• ScriptLine: accept empty text as "model produced no transcript" signal #520 (@wilke0818)
• Feat/pii subprocess venv #519 (@wilke0818)
• subprocess_venv: route torch/torchaudio via CUDA-aware PyTorch index #516 (@satra @wilke0818)
• subprocess_venv: two-stage install to fix torch/torchaudio CUDA-tag s… #518 (@wilke0818)
• Fix/ser wav2vec2 regression head #511 (@satra @wilke0818 @claude)
• comparison & uncertainty stage for analyze_audio.py #512 (@satra)
• audio analysis script + multilingual MMS aligner + 4 new ASR backends #510 (@satra)

Authors: 3

• Claude (@claude)
• Jordan Wilke (@wilke0818)
• Satrajit Ghosh (@satra)