Add click, typing and action sound effects mixed under the voice

Problem

The polished video now has a cursor, ripples and automatic zoom, but no interaction sounds. Demos need a coherent sound palette that supports the actions while keeping narration clear (story decisions 012–013; ticket decisions 001, 003–008; decision 002 is superseded).

Change

• Automatic short click sounds and quieter per-character typing taps, with three similar typing variants selected in a deterministic sequence.
• A synchronous sfx(name): void helper in the demo body context for eight explicit action cues: success, error, notification, open, close, toggle, back, select. Authors call it immediately after the relevant action or assertion; no UI heuristics.
• Use a selected subset of the web-sourced CC0 audition libraries (Kenney initially; DAEDAL after decision 008 listening feedback). Preserve original licence files, source revision, hashes and processing/gain provenance. Bundle selected runtime assets for offline execution; the broader audition library stays in planning.
• SFX are on by default. flow-reel record demo.ts --no-sfx disables every effect while preserving cue logging, narration, visuals and timing.
• Use the calibrated, held delivery timeline for every effect. Explicit cues suppress nearby automatic click audio, without removing clicks or ripples from the event log/video.
• A restrained, narration-first mix: quiet clicks/action cues, quieter typing; an additional 6 dB SFX duck over narration clips with 30 ms fade-down and 250 ms recovery. Preserve the existing voiced-delivery target of -16 LUFS and -1.5 dBTP peak ceiling. Silent-narration demos retain quiet fixed-level effects instead of being amplified to speech loudness.
• Effects never pause the script or extend video duration; tails at the delivery end are trimmed. Original speech WAVs, their hashes, the clean master, narration holds and captions retain their existing contracts.

Out of scope

Background music; automatic inference of success/error/navigation; custom sound filenames, gain controls or sound themes; additional automatic hover/scroll sounds; cloud TTS; unrelated recorder/CI fixes.

Done when

• A narrated fixture demonstrates automatic clicks/typing and every explicit cue; click onset is within one video frame of its ripple unless deliberately replaced by a nearby explicit cue.
• Cue calls neither wait nor change action/step timing. Old scripts work without modification, and older event logs remain readable by the updated code.
• Tests prove Unicode typing scheduling, deterministic variant selection, hold remapping, exact overlap boundaries, duck ramps, tail trimming, no-event/SFX-only paths and complete muting with --no-sfx.
• Measured SFX levels and narration intelligibility pass the calibration and decoded-media checks in impl.md; voiced delivery meets the inherited loudness target. Listening approval confirms sound selection and typing comfort beneath real narration.
• Identical timeline/events/assets produce identical decoded audio within the same recording image. Two browser runs retain the inherited transcript/source-audio reproducibility checks; this does not claim identical browser capture or MP4 bytes.
• Runtime assets and licences work from the packed npm package and recording image with no runtime downloads or access to the planning tree.

Evidence and plan

• Implementation plan
• Audition library and source/licence evidence
• Sound browser