recording: drive all tutorial recordings in content-time (lockstep follow-up)

[borisbat]

Follow-up to #191 (content-time helpers). Refactors all 47 tests/integration/record_*.das drivers off wall-clock sleep() pacing and onto the shared content-time helpers, then re-records every tutorial video.

Why

Under with_recording_app the host runs a fixed-dt clock — every captured frame is exactly 1/fps of video. A wall-clock sleep(ms) therefore stretches a beat by the render/playback fps ratio and varies with per-beat render load, so beats land unevenly and the playback judders. #191 added the content-time helpers (hold_content, record_frame_count, hold_remainder_content, *_voice action helpers) and moved the voice helpers over; this PR moves the remaining inline dwell/settle/warmup pacing in the drivers themselves.

What changed

• Pure settle/warmup/dwell drivers — sleep(...) → hold_content(app, ...) (gates on captured-frame count, falls back to wall-clock when not recording).
• Lead + remainder beats — restructured to anchor on record_frame_count(app), hold a short lead, run the action + checks, then hold_remainder_content(app, dwell, anchor) pads the beat to the voice dwell measuring consumed frames (dropped the old magic +Nu fudge constants).
• Centralized repeated patterns into *_voice helpers (e.g. sort_under_voice, context_pick_voice, open_under_voice/pick_under_voice, type_field_voice, expand_under_voice, drag_under_voice) so prepare_recording can still AST-scan the literal say lines at each call site.
• record_child / record_with_tab_stop / record_with_disabled — converted the remaining ref_time_ticks-based anchors to record_frame_count, and the special max(hold, body) dwell in with_disabled stage 4 to lead + hold_remainder_content.
• Re-recorded all 47 tutorial .mp4s on the content-time drivers (verified 47/47 play, correct durations, no dropped frames / overruns).
• record_group.das — dropped an unused require math (lint STYLE030).

move_to is already non-blocking (posts a lerped timeline consumed under the lockstep clock during the following hold), so only bare sleep() was a wall-clock leak.

Verification

• Compile 47/47, MCP lint clean (0 issues), formatter --verify clean (47 files) — all via the pre-push hook.
• Full re-record: 47/47 ok, 0 dropped, 0 overruns.
• Videos eyeballed and approved.

🤖 Generated with Claude Code