[codex] v3.28 follow-up: fix position hysteresis and voice cleanup

[Anthonyiswhy]

What changed

• makes get_position() frame-stable so repeated same-frame reads do not consume the left/right/ahead hysteresis budget
• threads frame_tag and frame_width through threat scoring so side-pass boosts use the same cached zone as the rest of the frame
• removes the unused dist_m parameter from _select_voice_message() so the helper matches its real interface
• makes PiperVoice.synthesize_alert_to_file() honor the instance _alert_cache_enabled flag instead of ignoring it
• cleans up empty temp WAVs if _copy_to_temp() fails mid-copy
• syncs repo metadata and interview notes to the current 174 passed validation count

Why

Claude's review surfaced a real field-impacting bug: get_position() mutates hysteresis state, but it was being called 4-6 times in one video frame. That meant objects near a zone boundary could switch from left/right to ahead faster than intended.

The other changes are cleanup and correctness fixes around the alert helper path so the code matches the behavior the tests and runtime expect.

Validation

• python -m py_compile raspberry_pi/yolo_realsense_navigation.py tests/test_blindnav_v326.py
• pytest tests/test_blindnav.py tests/test_blindnav_v326.py -q
• Result: 174 passed