Add push-to-talk mode and allow longer time to read problems

[Nomask-CHY]

Summary

During the interview, two conversational pacing issues degrade the practice experience:

1. Aggressive Turn-Taking on Slower Speech: For candidates who speak at a deliberate pace or pause briefly to construct sentences, the VAD (Voice Activity Detection) prematurely marks the end of an utterance, causing Jim to cut in before the sentence is finished.
2. Hardcoded / Overly Brief Reading Buffer: When asking for time to read the problem description or think through the approach, Jim often grants only a very short, rigid window (around 1 minute) and automatically interrupts/prompts the candidate even if more time was requested.

---

Motivation & Observed Behavior

• VAD Truncation: During complex explanations (e.g., REACTO Approach/Code phases), a 1~2 second pause does not mean the candidate has yielded the floor. Premature interruption fragments the candidate's line of thought and produces incomplete transcriptions.
• Reading Time Constraints: Comprehending algorithmic problem descriptions, constraints, and edge cases often requires 2–3 minutes. Fixed-interval prompting causes unnecessary pressure and disrupts the candidate's reading flow.

(Related to #140, #172, and #188)

---

Proposed Solutions

1. Push-to-Talk (PTT) Mode (Alternative to Auto-VAD)

• Add a configurable Push-to-Talk (PTT) toggle (e.g., hold Spacebar or a dedicated on-screen button to record, release to send).
• While Jim is speaking, disable recording input to prevent acoustic bleed and unintentional interruption.
• This provides an explicit boundary for utterances and completely eliminates false pauses caused by deliberate speech.

2. Configurable / Adaptive Silence Duration & Explicit Wait State

• Silence Threshold Tuning: Expose a setting or increase the default trailing silence detection window (e.g., debounce threshold from ~700ms to 1.5s–2.0s) before Jim claims the floor.
• Dedicated Reading State: When entering the problem inspection phase or when the candidate explicitly requests time (e.g., "Give me 2 minutes to read"), pause Jim's automated nudge timer until the candidate explicitly triggers an interaction (via voice or a "Ready" button).

---

Environment

• OS: Windows 11 (x86_64)
• Browser: Google Chrome / Microsoft Edge

[jserv]

Both parts of this were addressed by PR #192, which closed #140 and #172 and merged on 2026-10-01, after the 417b2c2 build many people are still running. It raised the default silence window from one second to three, added "Your turn is done" (a button, Alt+Enter, or a click on the countdown ring) to hand Jim the turn the moment you finish, and added a thinking hold: the Thinking button or a spoken "let me think" keeps the floor until you speak again or press Continue, with one check-in after two silent minutes. Please retest on the current latest release and add the output of codetrial --version. If the hold and the yield cover both cases, this will be closed. If you still want push-to-talk on top of them, narrow this issue to that and say which case the yield button leaves unsolved, since push-to-talk and the yield end the turn the same way.