Debug with an explicit hypothesis-and-evidence trail

Help me debug this issue using an explicit reasoning trail, so I can follow and check your logic instead of trusting a guess.

Symptom: [WHAT HAPPENS, INCLUDING EXACT ERROR MESSAGES OR WRONG OUTPUT]
Expected: [WHAT SHOULD HAPPEN]
When it started / what changed recently: [RECENT CHANGES, OR "UNKNOWN"]
Environment: [LANGUAGE/RUNTIME VERSIONS, OS, RELEVANT DEPENDENCIES]
Code and logs: [PASTE THE RELEVANT CODE AND ANY STACK TRACE OR LOG OUTPUT]

Work in this loop, and show each step:
1. Observations: list only facts established by the material above. No interpretation yet.
2. Hypotheses: enumerate at least three distinct causes consistent with all observations. For each, state what current evidence supports it and what evidence would rule it out.
3. Ranking: order the hypotheses by likelihood and say why — based on this evidence, not on which bug is most common in general.
4. Next probe: give me the single cheapest test that best separates the top hypotheses — a specific log line to add, value to print, input to try, or command to run. State the expected result under each hypothesis, so the outcome is informative either way.

I will run the probe and reply with the result. When you update your ranking, say explicitly which hypothesis was weakened or eliminated and why. Do not propose a fix until one hypothesis explains every observation, including the timing of when the bug appeared. If my reported result contradicts your leading hypothesis, treat the report as correct and revise.