recording: pace voice helpers in content-time (lockstep follow-up)

[borisbat]

Follow-up to the lockstep-recording sweep (#189). #189 converted say() to content-time and flagged the manual-hold drivers as a mechanical follow-up; this is that sweep, done centrally in the playwright helpers so every driver (here, and in the dependent implot / node-editor repos that call these same helpers) benefits at once.

Why

Under lockstep the recorder captures every rendered frame at a fixed 1/fps delay. A wall-clock sleep(ms) therefore yields render_rate/fps * ms of video — a 2 s caption shows for ~4 s on a 60 fps-rendering machine — and drifts the caption off its (real-time) voiceover. Content-time holds wait for ms/1000*fps captured frames, so the beat is exactly ms of video and caption / voice / visual stay aligned in the mux.

Changes (widgets/imgui_playwright.das)

• 18 pacing sleep(ms) → hold_content(app, ms) across hold_through_voice, drag_through_voice, type_into_voice, combo_pick_voice, toggle_tree_voice, close_button_voice, click_render_voice, menu_pick_voice. hold_content == sleep outside a recording session, so non-recording runs are unchanged. The synth mouse/key timelines ride the same fixed clock under lockstep, so the drag/click/type waits (e.g. hold_content(app, play)) stay in sync with the timeline they wait on.
• 6 helpers measured the consumed dwell in wall-clock (get_time_usec) then padded a content-time remainder — a mode mismatch that over-padded the caption tail 2–4× under recording. They now anchor in captured frames (record_frame_count) and pad via hold_remainder_content(app, dwell, started_frames).
• Comments updated (sleep → hold); the PACING-vs-POLL distinction is unchanged — pacing beats are content-time holds, effect-waits stay polled (wait_* / record_check_*).

Poll-machinery sleeps (READY/FRAME/RECORD poll intervals) are intentionally left as wall-clock.

Validation

No test impact — no test_*.das uses the voice helpers (recording-driver-only). Re-recorded edit_external_tour (drag + combo + type + hold, voiced, self-verifying):

exit 0   (all record_check_* pass -> content-time waits match the synth timelines)
dropped 0
2726 frames / 30 fps = 90.867 s exactly

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR updates the Playwright recording “voice helper” pacing so that, under lockstep recording, beats are held in content-time (captured-frame count) rather than wall-clock time, keeping captions/actions aligned with the recorded video timeline.

Changes:

• Replaced multiple pacing sleep(ms) calls in voice helpers with hold_content(app, ms) to hold by captured frames during recording.
• Switched several helpers from wall-clock “consumed dwell” tracking (get_time_usec) to frame-based anchoring (record_frame_count) and padding via hold_remainder_content.
• Updated helper comments to reflect the new hold-based pacing semantics.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.