Reels: burst beats + duration budget for week/month, plus step logging

Restructures a reel around beats — one narration line over one or more
photos — instead of one line per photo. A single-photo beat is a held
shot; a multi-photo beat is a quick burst that flashes through several
moments of an event while the line is read. So a week/month reel can show
everything it spans without a narrated (and timed) segment per photo.

Selection (selector.rs):
- Duration budget: cap the number of narrated beats to ~REEL_TARGET_SECONDS
  (default 90, env-tunable) so week/month reels don't run minutes long.
- Event clustering by time gap; when there are more events than the beat
  budget, adjacent events merge so the whole span stays covered. Each beat
  bursts up to MAX_BURST_PHOTOS (an even spread), so a 40-shot dinner
  contributes a handful of quick frames, not forty narrated seconds.

Render (render.rs): a beat renders its photos as a concat of per-photo
fills (blurred-bg portrait, fps-before-fade) under one muxed narration;
burst photos get a snappier fade. beat_durations splits the narration
across the photos, stretching only if a long burst would flash too fast.

Adds high-level info logs across the steps (request → script → per-beat
narrate/render → join → done with elapsed) for visibility. Bumps
RENDER_VERSION to re-render cached reels.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

src/reels/mod.rs

@@ -36,24 +36,27 @@ use crate::otel::extract_context_from_request;
 use crate::state::AppState;
 use selector::ReelSelector;
 
-/// The media behind one reel segment. Photos-only for now; a `Clip` variant
+/// The media behind one shot. Photos-only for now; a `Clip` variant (a section
-/// (a section of a source video) is the phase-2 extension point.
+/// of a source video) is the phase-2 extension point.
 #[derive(Debug, Clone)]
 pub enum SegmentMedia {
     Photo { rel_path: String, library_id: i32 },
 }
 
-/// A segment before narration: which photo, when it was taken, and any cached
+/// A beat: one narration line over one or more photos. A single-photo beat is a
-/// insight to feed the scripter.
+/// held shot; a multi-photo beat is a quick burst that flashes through several
+/// moments of the same event while the line is read — so a week/month reel can
+/// *show* everything it spans without a narration line (and the seconds that
+/// come with it) per photo.
 #[derive(Debug, Clone)]
-pub struct PlannedSegment {
+pub struct PlannedBeat {
-    pub media: SegmentMedia,
+    pub photos: Vec<SegmentMedia>,
     pub date: Option<i64>,
     pub insight_title: Option<String>,
     pub insight_summary: Option<String>,
 }
 
-impl PlannedSegment {
+impl PlannedBeat {
     /// Human date for the prompt, e.g. "June 12, 2019". `None` when undated.
     pub fn date_label(&self) -> Option<String> {
         let ts = self.date?;
@@ -180,7 +183,7 @@ fn finish_job(
 
 /// Render version: bump to invalidate every cached reel after a rendering /
 /// scripting change that should produce a fresh result.
-const RENDER_VERSION: u32 = 3;
+const RENDER_VERSION: u32 = 4;
 
 /// Narration expressiveness — Chatterbox's `exaggeration` knob. A slight bump
 /// over the ~0.5 default warms up otherwise-flat narration without over-acting;
@@ -306,16 +309,25 @@ pub async fn create_reel_handler(
         }));
     }
 
-    let media: Vec<SegmentMedia> = planned.iter().map(|p| p.media.clone()).collect();
+    // Flatten every photo across beats (in order) into the cache key — the key
+    // tracks exactly which photos appear and in what sequence.
+    let media: Vec<SegmentMedia> = planned.iter().flat_map(|b| b.photos.clone()).collect();
     let voice = req.voice.clone().filter(|s| !s.is_empty());
     let key = cache_key(&selector, &media, voice.as_deref());
 
     let job_id = Uuid::new_v4();
+    log::info!(
+        "reel {job_id}: request span={:?} → {} beats, {} photos",
+        span,
+        planned.len(),
+        media.len()
+    );
 
     // Cache hit: register an already-Done job pointing at the existing MP4 so
     // the client's first poll returns the video URL immediately.
     let mp4 = reel_mp4_path(&app_state, &key);
     if mp4.exists() {
+        log::info!("reel {job_id}: cache hit, serving existing reel");
         let title = std::fs::read(reel_sidecar_path(&app_state, &key))
             .ok()
             .and_then(|b| serde_json::from_slice::<ReelSidecar>(&b).ok())
@@ -358,6 +370,7 @@ pub async fn create_reel_handler(
             },
         );
     }
+    log::info!("reel {job_id}: queued for generation");
 
     let state = app_state.clone();
     let insight_dao = insight_dao.clone();
@@ -441,45 +454,73 @@ async fn run_reel_job(
     app_state: &AppState,
     insight_dao: &Mutex<Box<dyn InsightDao>>,
     job_id: Uuid,
-    mut planned: Vec<PlannedSegment>,
+    mut planned: Vec<PlannedBeat>,
     meta: ReelMeta,
     voice: Option<String>,
     key: &str,
 ) -> anyhow::Result<(String, PathBuf)> {
     use anyhow::{Context, anyhow};
 
+    let started = Instant::now();
+    let total_photos: usize = planned.iter().map(|b| b.photos.len()).sum();
+    log::info!(
+        "reel {job_id}: starting — span {:?}, {} beats, {} photos, voice={}",
+        meta.span,
+        planned.len(),
+        total_photos,
+        voice.as_deref().unwrap_or("default")
+    );
+
     let client = app_state
         .llamacpp
         .as_ref()
         .ok_or_else(|| anyhow!("TTS/LLM backend not configured"))?
         .clone();
 
-    // 1. Enrich with cached insights, then script (one LLM call).
+    // 1. Enrich each beat with its lead photo's cached insight, then script
+    //    (one LLM call → one narration line per beat).
     set_stage(job_id, "scripting");
+    log::info!("reel {job_id}: scripting narration via LLM…");
     let span_context = opentelemetry::Context::new();
     selector::enrich(insight_dao, &span_context, &mut planned);
     let script = script::generate_script(&client, &meta, &planned).await?;
+    log::info!(
+        "reel {job_id}: scripted \"{}\" ({} lines)",
+        script.title,
+        script.lines.len()
+    );
 
-    // 2. Narrate each line to speech and 3. render each photo segment. A
+    // 2. Narrate each beat's line and 3. render the beat (its photos shown in
-    // segment whose audio or render fails is skipped (logged) rather than
+    //    sequence under that one narration). A beat whose audio or render fails
-    // sinking the whole reel — handles an odd HEIC/corrupt file gracefully.
+    //    is skipped (logged) rather than sinking the whole reel — handles an
+    //    odd HEIC/corrupt file gracefully.
     set_stage(job_id, "narrating");
     let work = tempfile::tempdir().context("creating reel work dir")?;
     let nvenc = render::is_nvenc_available().await;
+    log::info!(
+        "reel {job_id}: narrating + rendering {} beats (encoder: {})",
+        planned.len(),
+        if nvenc { "nvenc" } else { "cpu" }
+    );
     let opts = render::SegmentOpts {
         nvenc,
         ..Default::default()
     };
 
-    let mut segment_files: Vec<String> = Vec::new();
+    let beat_total = planned.len();
-    for (i, (seg, line)) in planned.iter().zip(script.lines.iter()).enumerate() {
+    let mut beat_files: Vec<String> = Vec::new();
-        let image_path = match resolve_image_path(app_state, &seg.media) {
+    for (i, (beat, line)) in planned.iter().zip(script.lines.iter()).enumerate() {
-            Some(p) => p,
+        // Resolve all of the beat's photos to absolute paths; drop any that
-            None => {
+        // don't resolve. An empty beat is skipped.
-                log::warn!("reel {job_id}: skipping segment {i}, image path unresolved");
+        let image_paths: Vec<PathBuf> = beat
-                continue;
+            .photos
-            }
+            .iter()
-        };
+            .filter_map(|m| resolve_image_path(app_state, m))
+            .collect();
+        if image_paths.is_empty() {
+            log::warn!("reel {job_id}: skipping beat {i}, no image paths resolved");
+            continue;
+        }
 
         let audio_bytes = match crate::ai::tts::synthesize_serialized(
             &client,
@@ -492,13 +533,13 @@ async fn run_reel_job(
         {
             Ok(b) => b,
             Err(e) => {
-                log::warn!("reel {job_id}: skipping segment {i}, TTS failed: {e}");
+                log::warn!("reel {job_id}: skipping beat {i}, TTS failed: {e}");
                 continue;
             }
         };
         let audio_path = work.path().join(format!("narration_{i:03}.wav"));
         if let Err(e) = tokio::fs::write(&audio_path, &audio_bytes).await {
-            log::warn!("reel {job_id}: skipping segment {i}, writing audio failed: {e}");
+            log::warn!("reel {job_id}: skipping beat {i}, writing audio failed: {e}");
             continue;
         }
 
@@ -508,25 +549,37 @@ async fn run_reel_job(
                 .ok()
                 .flatten()
                 .unwrap_or(render::MIN_SEGMENT_SECONDS);
-        let duration = render::segment_duration(narration_secs);
 
         set_stage(job_id, "rendering");
-        let seg_out = work.path().join(format!("seg_{i:03}.mp4"));
+        log::info!(
+            "reel {job_id}: beat {}/{} — {} photo(s), narration {:.1}s",
+            i + 1,
+            beat_total,
+            image_paths.len(),
+            narration_secs
+        );
+        let beat_out = work.path().join(format!("beat_{i:03}.mp4"));
         if let Err(e) =
-            render::render_segment(&image_path, &audio_path, &seg_out, duration, &opts).await
+            render::render_beat(&image_paths, &audio_path, &beat_out, narration_secs, &opts).await
         {
-            log::warn!("reel {job_id}: skipping segment {i}, render failed: {e}");
+            log::warn!("reel {job_id}: skipping beat {i}, render failed: {e}");
             continue;
         }
-        segment_files.push(seg_out.to_string_lossy().to_string());
+        beat_files.push(beat_out.to_string_lossy().to_string());
     }
 
+    let segment_files = beat_files;
     if segment_files.is_empty() {
-        return Err(anyhow!("no segments rendered successfully"));
+        return Err(anyhow!("no beats rendered successfully"));
     }
 
     // 4. Concat into the cache. Write to a temp name in the reels dir, then
     // rename atomically (same filesystem) so a reader never sees a partial.
+    set_stage(job_id, "rendering");
+    log::info!(
+        "reel {job_id}: joining {} rendered beats into the final reel",
+        segment_files.len()
+    );
     std::fs::create_dir_all(&app_state.reels_path).context("creating reels dir")?;
     let final_path = reel_mp4_path(app_state, key);
     let tmp_path = final_path.with_extension("mp4.tmp");
@@ -541,6 +594,12 @@ async fn run_reel_job(
     .context("serializing reel sidecar")?;
     let _ = std::fs::write(reel_sidecar_path(app_state, key), sidecar);
 
+    log::info!(
+        "reel {job_id}: done in {:.1}s — {} beats → {}",
+        started.elapsed().as_secs_f64(),
+        segment_files.len(),
+        final_path.display()
+    );
     Ok((script.title, final_path))
 }
 
@@ -622,16 +681,16 @@ mod tests {
 
     #[test]
     fn date_label_formats_or_none() {
-        let seg = PlannedSegment {
+        let beat = PlannedBeat {
-            media: photo("a.jpg", 1),
+            photos: vec![photo("a.jpg", 1)],
             date: Some(1_560_384_000), // 2019-06-13 UTC
             insight_title: None,
             insight_summary: None,
         };
-        assert!(seg.date_label().unwrap().contains("2019"));
+        assert!(beat.date_label().unwrap().contains("2019"));
 
-        let undated = PlannedSegment {
+        let undated = PlannedBeat {
-            media: photo("a.jpg", 1),
+            photos: vec![photo("a.jpg", 1)],
             date: None,
             insight_title: None,
             insight_summary: None,

src/reels/render.rs

@@ -22,25 +22,31 @@ pub use crate::video::ffmpeg::is_nvenc_available;
 /// Reel canvas. Portrait, because reels are watched on a phone held upright —
 /// a landscape canvas letterboxes to a thin ~25%-height band there. Each photo
 /// is fitted sharp and centered over a blurred, zoomed copy of itself (see
-/// [`segment_filtergraph`]) so the frame is always filled regardless of the
+/// [`photo_filter_chain`]) so the frame is always filled regardless of the
 /// photo's orientation, without cropping the subject.
 pub const REEL_WIDTH: u32 = 1080;
 pub const REEL_HEIGHT: u32 = 1920;
 pub const REEL_FPS: u32 = 30;
 
-/// A still's screen time is its narration length plus a short breath, with a
+/// A beat's screen time is its narration length plus a short breath, with a
-/// floor so a terse line still lingers. No ceiling: the segment always covers
+/// floor so a terse line still lingers. No ceiling: the beat always covers the
-/// the full narration so speech is never truncated — the scripter is asked to
+/// full narration so speech is never truncated — the scripter is asked to keep
-/// keep lines short instead.
+/// lines short instead.
 pub const MIN_SEGMENT_SECONDS: f64 = 2.5;
 const NARRATION_TAIL_SECONDS: f64 = 0.6;
 
-/// Quick fade in/out baked into each segment so concatenated photos dip
+/// Fade durations baked into each photo. A held (single-photo) beat gets a
-/// smoothly instead of hard-cutting. The fade-out lands inside the narration's
+/// gentle dip; burst photos get a snappier fade so the montage feels quick.
-/// silent tail, so speech is never clipped.
+const SINGLE_FADE_SECONDS: f64 = 0.35;
-const FADE_SECONDS: f64 = 0.35;
+const BURST_FADE_SECONDS: f64 = 0.15;
 
-/// Screen time for a photo segment given its narration audio length.
+/// Floor on how long each burst photo stays up, so a long line over many photos
+/// doesn't flash them subliminally. If the narration is too short to give every
+/// photo this much, the beat is stretched to fit.
+const MIN_BURST_PHOTO_SECONDS: f64 = 0.6;
+
+/// Base screen time for a beat given its narration length: narration + breath,
+/// floored. Used as the lower bound on a beat's total duration.
 pub fn segment_duration(narration_secs: f64) -> f64 {
     let d = narration_secs + NARRATION_TAIL_SECONDS;
     if d.is_finite() && d > MIN_SEGMENT_SECONDS {
@@ -50,6 +56,29 @@ pub fn segment_duration(narration_secs: f64) -> f64 {
     }
 }
 
+/// Split a beat into per-photo durations. The beat lasts at least its narration
+/// (so speech isn't cut) and at least `n × MIN_BURST_PHOTO_SECONDS` (so a fast
+/// burst stays legible); the photos share that total evenly. Returns
+/// `(total_seconds, per_photo_seconds)`.
+pub fn beat_durations(narration_secs: f64, n_photos: usize) -> (f64, Vec<f64>) {
+    let n = n_photos.max(1);
+    let base = segment_duration(narration_secs);
+    let min_total = n as f64 * MIN_BURST_PHOTO_SECONDS;
+    let total = if base > min_total { base } else { min_total };
+    let each = total / n as f64;
+    (total, vec![each; n])
+}
+
+/// Fade length to use for a beat of `n_photos` (gentle when held, snappy in a
+/// burst).
+fn fade_for(n_photos: usize) -> f64 {
+    if n_photos > 1 {
+        BURST_FADE_SECONDS
+    } else {
+        SINGLE_FADE_SECONDS
+    }
+}
+
 /// Options controlling per-segment rendering.
 #[derive(Debug, Clone, Copy)]
 pub struct SegmentOpts {
@@ -70,38 +99,49 @@ impl Default for SegmentOpts {
     }
 }
 
-/// Full `filter_complex` for one photo segment, producing labelled `[v]` (video)
+/// Filter chain for one photo (input `idx`) producing the labelled output
-/// and `[a]` (audio) outputs. Input 0 is the looped still, input 1 the
+/// `[v{idx}]`. Splits the still into a background and foreground: the background
-/// narration.
+/// is scaled to *cover* the canvas and heavily blurred; the foreground is
+/// scaled to *fit* and overlaid centered. This fills the portrait frame for any
+/// photo orientation — no black bars, no cropping of the subject — then a fade
+/// in/out softens the cut. Intermediate labels are suffixed with `idx` so
+/// several chains coexist in one `filter_complex`.
 ///
-/// Video: split the still into a background and foreground. The background is
+/// `fps` is normalized BEFORE the fades so the brightness ramp is computed on a
-/// scaled to *cover* the canvas and heavily blurred; the foreground is scaled to
+/// true {fps}-frame timeline; otherwise the fade is sampled at the looped
-/// *fit* inside it and overlaid centered. This fills the portrait frame for any
+/// still's coarse cadence and duplicated up, which reads as a steppy dip.
-/// photo orientation — no black bars, no cropping of the subject — then a quick
+fn photo_filter_chain(idx: usize, opts: &SegmentOpts, duration: f64, fade: f64) -> String {
-/// fade in/out softens the cut to the next segment.
-///
-/// Audio: pad the narration with trailing silence so a short line doesn't end
-/// the segment early; `-t` bounds it to the segment duration.
-pub fn segment_filtergraph(opts: &SegmentOpts, duration: f64) -> String {
     let (w, h, fps) = (opts.width, opts.height, opts.fps);
-    // Fade-out begins one fade-length before the end; clamp so a floor-length
+    let fade_out_start = (duration - fade).max(0.0);
-    // segment still gets a valid (non-negative) start time.
-    let fade_out_start = (duration - FADE_SECONDS).max(0.0);
-    // `fps` is normalized BEFORE the fades so the brightness ramp is computed
-    // on a true {fps}-frame timeline. If fps came after, the fade would be
-    // sampled at the looped still's coarse input cadence and then duplicated up
-    // to {fps}, which reads as a steppy / low-frame-rate dip.
     format!(
-        "[0:v]split=2[bg][fg];\
+        "[{idx}:v]split=2[bg{idx}][fg{idx}];\
-         [bg]scale={w}:{h}:force_original_aspect_ratio=increase,\
+         [bg{idx}]scale={w}:{h}:force_original_aspect_ratio=increase,\
-         crop={w}:{h},boxblur=20:2[bgb];\
+         crop={w}:{h},boxblur=20:2[bgb{idx}];\
-         [fg]scale={w}:{h}:force_original_aspect_ratio=decrease[fgs];\
+         [fg{idx}]scale={w}:{h}:force_original_aspect_ratio=decrease[fgs{idx}];\
-         [bgb][fgs]overlay=(W-w)/2:(H-h)/2,\
+         [bgb{idx}][fgs{idx}]overlay=(W-w)/2:(H-h)/2,\
          fps={fps},\
-         fade=t=in:st=0:d={FADE_SECONDS},\
+         fade=t=in:st=0:d={fade},\
-         fade=t=out:st={fade_out_start:.3}:d={FADE_SECONDS},\
+         fade=t=out:st={fade_out_start:.3}:d={fade},\
-         setsar=1,format=yuv420p[v];\
+         setsar=1,format=yuv420p[v{idx}]"
-         [1:a]apad[a]"
+    )
+}
+
+/// Full `filter_complex` for a beat of `per_photo` durations: one chain per
+/// photo, concatenated into `[v]`, with the narration (the last input, index
+/// `per_photo.len()`) padded with trailing silence into `[a]`. A single-photo
+/// beat degenerates to one chain + `concat=n=1` (a passthrough).
+pub fn beat_filtergraph(opts: &SegmentOpts, per_photo: &[f64]) -> String {
+    let n = per_photo.len().max(1);
+    let fade = fade_for(n);
+    let chains: Vec<String> = per_photo
+        .iter()
+        .enumerate()
+        .map(|(i, &d)| photo_filter_chain(i, opts, d, fade))
+        .collect();
+    let concat_inputs: String = (0..n).map(|i| format!("[v{i}]")).collect();
+    format!(
+        "{chains};{concat_inputs}concat=n={n}:v=1:a=0[v];[{n}:a]apad[a]",
+        chains = chains.join(";")
     )
 }
 
@@ -128,15 +168,16 @@ fn video_encoder_args(nvenc: bool) -> Vec<String> {
     .collect()
 }
 
-/// Build the ffmpeg args that render one photo segment: a still looped for
+/// Build the ffmpeg args that render one beat: each photo looped for its slice
-/// `duration` seconds, filled to the portrait canvas with a blurred backdrop
+/// of the beat (filled to the portrait canvas with a blurred backdrop), the
-/// (see [`segment_filtergraph`]) and the narration muxed in. `-t` bounds both
+/// slices concatenated, and the single narration muxed over the whole thing.
-/// streams to the segment length.
+/// `total` bounds the output (and the apad'd audio) to the beat length.
-pub fn build_segment_args(
+pub fn build_beat_args(
-    image_path: &str,
+    image_paths: &[String],
     audio_path: &str,
     out_path: &str,
-    duration: f64,
+    per_photo: &[f64],
+    total: f64,
     opts: &SegmentOpts,
 ) -> Vec<String> {
     let fps = opts.fps.to_string();
@@ -144,26 +185,33 @@ pub fn build_segment_args(
     if opts.nvenc {
         args.extend(["-hwaccel".into(), "cuda".into()]);
     }
+    // One looped-still input per photo, each bounded to its slice by an input
+    // `-t`; reading at the target `-framerate` gives the fades real frames to
+    // ramp across.
+    for (path, &dur) in image_paths.iter().zip(per_photo.iter()) {
+        args.extend([
+            "-framerate".into(),
+            fps.clone(),
+            "-loop".into(),
+            "1".into(),
+            "-t".into(),
+            format!("{dur:.3}"),
+            "-i".into(),
+            path.clone(),
+        ]);
+    }
     args.extend([
-        // Read the looped still at the target rate so frames exist for the
-        // fade to ramp across (paired with the in-graph `fps` and CFR output).
-        "-framerate".into(),
-        fps.clone(),
-        "-loop".into(),
-        "1".into(),
-        "-i".into(),
-        image_path.into(),
         "-i".into(),
         audio_path.into(),
         "-filter_complex".into(),
-        segment_filtergraph(opts, duration),
+        beat_filtergraph(opts, per_photo),
         "-map".into(),
         "[v]".into(),
         "-map".into(),
         "[a]".into(),
         "-t".into(),
-        format!("{duration:.3}"),
+        format!("{total:.3}"),
-        // Force constant frame rate so the segment (and the concatenated reel)
+        // Force constant frame rate so the beat (and the concatenated reel)
         // plays at a steady {fps} rather than a variable cadence.
         "-r".into(),
         fps,
@@ -231,22 +279,33 @@ async fn run_ffmpeg(args: &[String], what: &str) -> Result<()> {
     Ok(())
 }
 
-/// Render one photo segment to `out_path`.
+/// Render one beat to `out_path`: its photos shown in sequence (a held shot for
-pub async fn render_segment(
+/// one photo, a quick burst for several) under the single narration in
-    image_path: &Path,
+/// `audio_path`, whose measured length sets the beat's pacing.
+pub async fn render_beat(
+    image_paths: &[std::path::PathBuf],
     audio_path: &Path,
     out_path: &Path,
-    duration: f64,
+    narration_secs: f64,
     opts: &SegmentOpts,
 ) -> Result<()> {
-    let args = build_segment_args(
+    if image_paths.is_empty() {
-        &image_path.to_string_lossy(),
+        bail!("render_beat called with no images");
+    }
+    let (total, per_photo) = beat_durations(narration_secs, image_paths.len());
+    let paths: Vec<String> = image_paths
+        .iter()
+        .map(|p| p.to_string_lossy().to_string())
+        .collect();
+    let args = build_beat_args(
+        &paths,
         &audio_path.to_string_lossy(),
         &out_path.to_string_lossy(),
-        duration,
+        &per_photo,
+        total,
         opts,
     );
-    run_ffmpeg(&args, "segment render").await
+    run_ffmpeg(&args, "beat render").await
 }
 
 /// Join rendered segments into the final reel. Writes the concat list into the
@@ -288,73 +347,108 @@ mod tests {
     }
 
     #[test]
-    fn filtergraph_fills_portrait_with_blurred_bg_and_fitted_fg() {
+    fn beat_durations_single_photo_matches_base() {
-        let g = segment_filtergraph(&SegmentOpts::default(), 4.0);
+        let (total, per) = beat_durations(4.0, 1);
-        // Background covers + blurs; foreground fits and is centered over it.
+        assert!((total - 4.6).abs() < 1e-9); // narration + tail
-        assert!(g.contains("split=2[bg][fg]"));
+        assert_eq!(per.len(), 1);
+        assert!((per[0] - 4.6).abs() < 1e-9);
+    }
+
+    #[test]
+    fn beat_durations_burst_splits_evenly() {
+        // 5 photos, narration 4.6s base → ~0.92s each (above the 0.6 floor).
+        let (total, per) = beat_durations(4.0, 5);
+        assert!((total - 4.6).abs() < 1e-9);
+        assert_eq!(per.len(), 5);
+        assert!((per.iter().sum::<f64>() - total).abs() < 1e-9);
+        assert!(per.iter().all(|&d| d >= MIN_BURST_PHOTO_SECONDS));
+    }
+
+    #[test]
+    fn beat_durations_stretches_when_narration_too_short_for_burst() {
+        // Floor narration (2.5s) over 10 photos would be 0.25s each — below the
+        // legibility floor, so the beat stretches to 10 × 0.6 = 6s.
+        let (total, per) = beat_durations(0.0, 10);
+        assert!((total - 6.0).abs() < 1e-9);
+        assert!(per.iter().all(|&d| (d - 0.6).abs() < 1e-9));
+    }
+
+    #[test]
+    fn beat_filtergraph_single_photo_fills_portrait_and_holds() {
+        let (_t, per) = beat_durations(4.0, 1);
+        let g = beat_filtergraph(&SegmentOpts::default(), &per);
+        assert!(g.contains("[0:v]split=2[bg0][fg0]"));
         assert!(g.contains("scale=1080:1920:force_original_aspect_ratio=increase"));
         assert!(g.contains("crop=1080:1920"));
-        assert!(g.contains("boxblur"));
         assert!(g.contains("scale=1080:1920:force_original_aspect_ratio=decrease"));
         assert!(g.contains("overlay=(W-w)/2:(H-h)/2"));
-        // Produces the labelled outputs build_segment_args maps.
+        // Single photo → concat of one, gentle fade, audio is input 1.
-        assert!(g.contains("[v]"));
+        assert!(g.contains("concat=n=1:v=1:a=0[v]"));
+        assert!(g.contains("d=0.35")); // SINGLE_FADE
         assert!(g.contains("[1:a]apad[a]"));
-        assert!(g.contains("format=yuv420p"));
     }
 
     #[test]
-    fn filtergraph_fades_in_and_out_within_duration() {
+    fn beat_filtergraph_burst_chains_concats_and_snappy_fade() {
-        // 4s segment, 0.35s fade → fade-out starts at 3.65s.
+        let (_t, per) = beat_durations(4.0, 3);
-        let g = segment_filtergraph(&SegmentOpts::default(), 4.0);
+        let g = beat_filtergraph(&SegmentOpts::default(), &per);
-        assert!(g.contains("fade=t=in:st=0:d=0.35"));
+        // One chain per photo with index-suffixed labels.
-        assert!(g.contains("fade=t=out:st=3.650:d=0.35"));
+        assert!(g.contains("[0:v]split") && g.contains("[1:v]split") && g.contains("[2:v]split"));
+        // Concatenated in order, audio is the 4th input (index 3).
+        assert!(g.contains("[v0][v1][v2]concat=n=3:v=1:a=0[v]"));
+        assert!(g.contains("[3:a]apad[a]"));
+        // Burst uses the snappier fade.
+        assert!(g.contains("d=0.15"));
+        assert!(!g.contains("d=0.35"));
     }
 
     #[test]
-    fn filtergraph_normalizes_fps_before_fading() {
+    fn beat_filtergraph_normalizes_fps_before_fading() {
-        // The fps filter must precede the fades, else the brightness ramp is
+        // fps must precede the fades on every chain (else the dip looks steppy).
-        // sampled at the still's coarse cadence and looks steppy.
+        let (_t, per) = beat_durations(4.0, 1);
-        let g = segment_filtergraph(&SegmentOpts::default(), 4.0);
+        let g = beat_filtergraph(&SegmentOpts::default(), &per);
         let fps_at = g.find("fps=30").expect("fps in graph");
         let fade_at = g.find("fade=t=in").expect("fade in graph");
         assert!(fps_at < fade_at);
     }
 
     #[test]
-    fn filtergraph_fade_out_start_never_negative_at_floor() {
+    fn beat_args_one_input_per_photo_plus_audio_bound_by_total() {
-        // A floor-length segment shorter than a fade still yields st >= 0.
+        let (total, per) = beat_durations(4.0, 2);
-        let g = segment_filtergraph(&SegmentOpts::default(), 0.2);
+        let args = build_beat_args(
-        assert!(g.contains("fade=t=out:st=0.000:d=0.35"));
+            &["/a.jpg".into(), "/b.jpg".into()],
-    }
+            "/n.wav",
-
-    #[test]
-    fn segment_args_loop_still_and_bound_with_t() {
-        let args = build_segment_args(
-            "/img.jpg",
-            "/a.wav",
             "/out.mp4",
-            4.0,
+            &per,
+            total,
             &SegmentOpts::default(),
         );
         let joined = args.join(" ");
-        assert!(joined.contains("-framerate 30 -loop 1 -i /img.jpg"));
+        // A looped-still input per photo, each with its slice -t, then the audio.
-        assert!(joined.contains("-i /a.wav"));
+        assert!(joined.contains("-framerate 30 -loop 1 -t 2.300 -i /a.jpg"));
-        assert!(joined.contains("apad"));
+        assert!(joined.contains("-framerate 30 -loop 1 -t 2.300 -i /b.jpg"));
-        assert!(joined.contains("-t 4.000"));
+        assert!(joined.contains("-i /n.wav"));
-        // Constant frame rate forced on the output.
+        // Output bounded to the beat total and forced CFR.
+        assert!(joined.contains("-t 4.600"));
         assert!(joined.contains("-r 30"));
-        assert!(joined.contains("libx264"));
         assert!(joined.ends_with("/out.mp4"));
     }
 
     #[test]
-    fn segment_args_use_nvenc_and_cuda_when_enabled() {
+    fn beat_args_use_nvenc_and_cuda_when_enabled() {
         let opts = SegmentOpts {
             nvenc: true,
             ..SegmentOpts::default()
         };
-        let args = build_segment_args("/img.jpg", "/a.wav", "/out.mp4", 3.0, &opts);
+        let (total, per) = beat_durations(3.0, 1);
+        let args = build_beat_args(
+            &["/img.jpg".into()],
+            "/a.wav",
+            "/out.mp4",
+            &per,
+            total,
+            &opts,
+        );
         let joined = args.join(" ");
         assert!(joined.contains("-hwaccel cuda"));
         assert!(joined.contains("h264_nvenc"));

src/reels/script.rs

@@ -1,10 +1,11 @@
 //! Narration scripting for memory reels.
 //!
-//! One LLM call turns the planned segments (each carrying its date and, where
+//! One LLM call turns the planned beats (each carrying its date and, where
 //! available, its cached insight) into a short first-person narration line per
-//! photo plus a title for the reel. We reuse the cached insight summary as the
+//! beat plus a title for the reel. A beat may show several photos in a quick
-//! richest per-photo signal rather than re-running vision at reel time — that
+//! burst, so a line narrates the *moment*, not a single frame. We reuse the
-//! keeps reel generation off the GPU's vision slot entirely.
+//! cached insight summary as the richest signal rather than re-running vision
+//! at reel time — that keeps reel generation off the GPU's vision slot.
 //!
 //! The prompt builder and response parser are pure so the contract is
 //! unit-testable; `generate_script` wires them to the LLM client.
@@ -12,11 +13,11 @@
 use anyhow::{Context, Result};
 use std::sync::Arc;
 
-use super::{PlannedSegment, ReelMeta};
+use super::{PlannedBeat, ReelMeta};
 use crate::ai::llamacpp::LlamaCppClient;
 use crate::ai::llm_client::LlmClient;
 
-/// The narration for a whole reel: a title and one line per segment, in order.
+/// The narration for a whole reel: a title and one line per beat, in order.
 #[derive(Debug, Clone, PartialEq)]
 pub struct ReelScript {
     pub title: String,
@@ -26,33 +27,38 @@ pub struct ReelScript {
 const SYSTEM_PROMPT: &str = "You are narrating a personal memory reel — a short \
 slideshow of someone's own photos set to a spoken voiceover. Write warm, \
 specific, first-person narration as if the person is gently looking back on \
-their own memories. Be concrete and grounded in the details given; never \
+their own memories. Each line plays over one moment, which may be a quick burst \
-invent names, places, or events that aren't supported. Keep each line to one \
+of several photos, so narrate the moment as a whole rather than a single frame. \
-or two short sentences that can be read aloud in a few seconds. Avoid generic \
+Be concrete and grounded in the details given; never invent names, places, or \
-filler like \"what a wonderful day\" — if you have little to go on, simply \
+events that aren't supported. Keep each line to one or two short sentences that \
-describe the moment plainly.";
+can be read aloud in a few seconds. Avoid generic filler like \"what a \
+wonderful day\" — if you have little to go on, simply describe the moment \
+plainly.";
 
 /// Build the (system, user) prompt pair for the scripter. The user message
-/// describes each segment in order and asks for strict JSON back.
+/// describes each beat in order and asks for strict JSON back.
-pub fn build_script_messages(meta: &ReelMeta, planned: &[PlannedSegment]) -> (String, String) {
+pub fn build_script_messages(meta: &ReelMeta, beats: &[PlannedBeat]) -> (String, String) {
     let mut user = String::new();
     user.push_str(&format!(
-        "These are {} photos surfaced as memories {}.\n\n",
+        "This reel has {} moments surfaced as memories {}.\n\n",
-        planned.len(),
+        beats.len(),
         meta.span_phrase()
     ));
     if !meta.years.is_empty() {
         let years: Vec<String> = meta.years.iter().map(|y| y.to_string()).collect();
         user.push_str(&format!("They span the years: {}.\n\n", years.join(", ")));
     }
-    user.push_str("Photos, in the order they will appear:\n");
+    user.push_str("Moments, in the order they will appear:\n");
-    for (i, seg) in planned.iter().enumerate() {
+    for (i, beat) in beats.iter().enumerate() {
         user.push_str(&format!("\n[{}]", i + 1));
-        if let Some(date) = seg.date_label() {
+        if let Some(date) = beat.date_label() {
             user.push_str(&format!(" {date}"));
         }
+        if beat.photos.len() > 1 {
+            user.push_str(&format!(" (a burst of {} photos)", beat.photos.len()));
+        }
         user.push('\n');
-        match (&seg.insight_title, &seg.insight_summary) {
+        match (&beat.insight_title, &beat.insight_summary) {
             (Some(t), Some(s)) if !s.trim().is_empty() => {
                 user.push_str(&format!("  Known context: {t} — {s}\n"));
             }
@@ -65,10 +71,10 @@ pub fn build_script_messages(meta: &ReelMeta, planned: &[PlannedSegment]) -> (St
     }
     user.push_str(&format!(
         "\nReturn ONLY a JSON object, no prose or code fences, shaped exactly:\n\
-         {{\"title\": \"<short reel title>\", \"segments\": [\"<line for photo 1>\", \
+         {{\"title\": \"<short reel title>\", \"segments\": [\"<line for moment 1>\", \
-         \"<line for photo 2>\", ... ]}}\n\
+         \"<line for moment 2>\", ... ]}}\n\
-         The \"segments\" array MUST have exactly {} items, one per photo in order.",
+         The \"segments\" array MUST have exactly {} items, one per moment in order.",
-        planned.len()
+        beats.len()
     ));
     (SYSTEM_PROMPT.to_string(), user)
 }
@@ -174,20 +180,20 @@ fn clean_text(s: &str) -> String {
     trimmed.split_whitespace().collect::<Vec<_>>().join(" ")
 }
 
-/// Generate the reel script via the LLM. Text-only (no images) — the per-photo
+/// Generate the reel script via the LLM. Text-only (no images) — the per-beat
 /// context comes from cached insights. The call takes the GPU read lease
 /// internally (see `LlamaCppClient::generate`).
 pub async fn generate_script(
     client: &Arc<LlamaCppClient>,
     meta: &ReelMeta,
-    planned: &[PlannedSegment],
+    beats: &[PlannedBeat],
 ) -> Result<ReelScript> {
-    let (system, user) = build_script_messages(meta, planned);
+    let (system, user) = build_script_messages(meta, beats);
     let raw = client
         .generate(&user, Some(&system), None)
         .await
         .context("LLM script generation failed")?;
-    Ok(parse_script_response(&raw, planned.len()))
+    Ok(parse_script_response(&raw, beats.len()))
 }
 
 #[cfg(test)]
@@ -202,13 +208,13 @@ mod tests {
         }
     }
 
-    fn planned(n: usize) -> Vec<PlannedSegment> {
+    fn planned(n: usize) -> Vec<PlannedBeat> {
         (0..n)
-            .map(|i| PlannedSegment {
+            .map(|i| PlannedBeat {
-                media: super::super::SegmentMedia::Photo {
+                photos: vec![super::super::SegmentMedia::Photo {
                     rel_path: format!("p{i}.jpg"),
                     library_id: 1,
-                },
+                }],
                 date: Some(1_560_000_000 + i as i64 * 86_400),
                 insight_title: None,
                 insight_summary: None,
@@ -217,16 +223,37 @@ mod tests {
     }
 
     #[test]
-    fn prompt_states_exact_segment_count_and_span() {
+    fn prompt_states_exact_moment_count_and_span() {
         let (sys, user) = build_script_messages(&meta(), &planned(3));
         assert!(sys.contains("memory reel"));
-        assert!(user.contains("3 photos"));
+        assert!(user.contains("3 moments"));
         assert!(user.contains("on this day"));
         assert!(user.contains("exactly 3 items"));
-        // Each photo gets an indexed entry.
+        // Each moment gets an indexed entry.
         assert!(user.contains("[1]") && user.contains("[2]") && user.contains("[3]"));
     }
 
+    #[test]
+    fn prompt_notes_burst_photo_count() {
+        let mut p = planned(1);
+        p[0].photos = vec![
+            super::super::SegmentMedia::Photo {
+                rel_path: "a.jpg".into(),
+                library_id: 1,
+            },
+            super::super::SegmentMedia::Photo {
+                rel_path: "b.jpg".into(),
+                library_id: 1,
+            },
+            super::super::SegmentMedia::Photo {
+                rel_path: "c.jpg".into(),
+                library_id: 1,
+            },
+        ];
+        let (_sys, user) = build_script_messages(&meta(), &p);
+        assert!(user.contains("a burst of 3 photos"));
+    }
+
     #[test]
     fn prompt_includes_insight_context_when_present() {
         let mut p = planned(1);

src/reels/selector.rs

@@ -13,18 +13,51 @@ use std::sync::Mutex;
 
 use chrono::{DateTime, Datelike, FixedOffset};
 
-use super::{PlannedSegment, ReelMeta, SegmentMedia};
+use super::{PlannedBeat, ReelMeta, SegmentMedia};
 use crate::database::{ExifDao, InsightDao};
 use crate::file_types::is_image_file;
 use crate::memories::{self, MemoriesSpan};
 use crate::state::AppState;
 
-/// Default and hard caps on how many photos a reel covers. The cap bounds the
+/// Default and hard caps on how many photos a reel covers. The default is an
-/// LLM/TTS/ffmpeg work per reel; when a span has more, [`sample_evenly`] keeps
+/// upper bound on the request; the effective count is usually smaller, set by
-/// a representative spread across the years rather than just the oldest.
+/// the duration budget (see [`budget_segments`]). The hard cap bounds work per
-pub const DEFAULT_MAX_SEGMENTS: usize = 24;
+/// reel regardless.
+pub const DEFAULT_MAX_SEGMENTS: usize = 40;
 pub const HARD_MAX_SEGMENTS: usize = 40;
 
+/// Target reel length. Week and especially month spans can surface hundreds of
+/// photos; at a few seconds of narration each, a naive reel runs minutes. We
+/// cap the segment count to keep the reel near this length. Tunable via
+/// `REEL_TARGET_SECONDS`.
+const DEFAULT_TARGET_REEL_SECONDS: f64 = 90.0;
+
+/// Rough average wall-time per photo segment (a short narration line + the
+/// silent tail). Only used to turn the duration target into a segment count;
+/// the real per-segment time is the measured narration length.
+const EST_SECONDS_PER_SEGMENT: f64 = 5.0;
+
+/// Time gap that separates one "event/moment" from the next when clustering a
+/// span's photos. Photos within a few hours are treated as the same occasion
+/// (and across years/days the gaps are far larger, so each instance clusters
+/// on its own). 4 hours splits e.g. a morning hike from an evening dinner.
+const EVENT_GAP_SECONDS: i64 = 4 * 3600;
+
+fn target_reel_seconds() -> f64 {
+    std::env::var("REEL_TARGET_SECONDS")
+        .ok()
+        .and_then(|s| s.trim().parse::<f64>().ok())
+        .filter(|x| x.is_finite() && *x > 0.0)
+        .unwrap_or(DEFAULT_TARGET_REEL_SECONDS)
+}
+
+/// How many photo segments fit the duration budget, bounded by the request's
+/// max and the hard cap. This is what keeps week/month reels from running long.
+pub fn budget_segments(requested_max: usize) -> usize {
+    let by_budget = (target_reel_seconds() / EST_SECONDS_PER_SEGMENT).floor() as usize;
+    by_budget.min(requested_max).clamp(1, HARD_MAX_SEGMENTS)
+}
+
 /// What a reel is built from. v1 ships the memories (on this day/week/month)
 /// selector; tag and date-range variants slot in here later.
 #[derive(Debug, Clone)]
@@ -81,6 +114,104 @@ pub fn sample_evenly<T: Clone>(items: &[T], max: usize) -> Vec<T> {
         .collect()
 }
 
+/// Group time-sorted items into events by gap: a new event starts whenever the
+/// jump from the previous photo exceeds `gap_seconds`. Preserves order; items
+/// without a timestamp extend the current event.
+fn cluster_by_gap(
+    items: &[memories::MemoryItem],
+    gap_seconds: i64,
+) -> Vec<Vec<memories::MemoryItem>> {
+    let mut clusters: Vec<Vec<memories::MemoryItem>> = Vec::new();
+    let mut prev_ts: Option<i64> = None;
+    for it in items {
+        let starts_new = match (prev_ts, it.created) {
+            (Some(p), Some(c)) => c - p > gap_seconds,
+            _ => false,
+        };
+        if starts_new || clusters.is_empty() {
+            clusters.push(Vec::new());
+        }
+        clusters.last_mut().unwrap().push(it.clone());
+        if let Some(c) = it.created {
+            prev_ts = Some(c);
+        }
+    }
+    clusters
+}
+
+/// Most photos a single beat will flash through. Bounds the burst so one huge
+/// event doesn't dominate, and keeps each photo on screen long enough to
+/// register at the per-beat narration length (see render's beat timing).
+pub const MAX_BURST_PHOTOS: usize = 10;
+
+/// Merge a list of (time-ordered) event clusters into exactly `n` contiguous
+/// groups, so a span with more events than the beat budget still covers the
+/// whole timeline — adjacent events fold together into one beat rather than
+/// getting dropped. `n` must be ≥ 1 and ≤ clusters.len().
+fn partition_into_groups(
+    clusters: Vec<Vec<memories::MemoryItem>>,
+    n: usize,
+) -> Vec<Vec<memories::MemoryItem>> {
+    let c = clusters.len();
+    let mut clusters = clusters.into_iter();
+    (0..n)
+        .map(|j| {
+            // Even contiguous split of c clusters into n groups.
+            let start = j * c / n;
+            let end = (j + 1) * c / n;
+            let take = end.saturating_sub(start).max(1);
+            (0..take)
+                .flat_map(|_| clusters.next().into_iter().flatten())
+                .collect()
+        })
+        .collect()
+}
+
+/// Turn a span's photos into `n_beats` beats. Clusters photos into events by
+/// time gap; if there are more events than beats, adjacent events are merged so
+/// the whole span is still covered. Each beat then flashes up to
+/// `max_burst` photos (an even spread of its group) under one narration line —
+/// so a week/month reel *shows* all its moments without a narrated (and timed)
+/// segment per photo.
+pub fn form_beats(
+    items: &[memories::MemoryItem],
+    n_beats: usize,
+    max_burst: usize,
+) -> Vec<PlannedBeat> {
+    if n_beats == 0 || items.is_empty() {
+        return Vec::new();
+    }
+    let clusters = cluster_by_gap(items, EVENT_GAP_SECONDS);
+    // One beat per event when they fit; otherwise fold adjacent events together
+    // into exactly n_beats groups.
+    let groups = if clusters.len() <= n_beats {
+        clusters
+    } else {
+        partition_into_groups(clusters, n_beats)
+    };
+
+    groups
+        .into_iter()
+        .filter(|g| !g.is_empty())
+        .map(|group| {
+            let shown = sample_evenly(&group, max_burst);
+            let date = shown.first().and_then(|it| it.created);
+            PlannedBeat {
+                photos: shown
+                    .into_iter()
+                    .map(|it| SegmentMedia::Photo {
+                        rel_path: it.path,
+                        library_id: it.library_id,
+                    })
+                    .collect(),
+                date,
+                insight_title: None,
+                insight_summary: None,
+            }
+        })
+        .collect()
+}
+
 /// Cheap pass: resolve the selector into an ordered list of media (no insight
 /// lookups yet) plus reel metadata. `Err` only on an invalid library param.
 pub fn resolve(
@@ -88,7 +219,7 @@ pub fn resolve(
     exif_dao: &Mutex<Box<dyn ExifDao>>,
     span_context: &opentelemetry::Context,
     selector: &ReelSelector,
-) -> Result<(Vec<PlannedSegment>, ReelMeta), String> {
+) -> Result<(Vec<PlannedBeat>, ReelMeta), String> {
     match selector {
         ReelSelector::Memories {
             span,
@@ -108,32 +239,23 @@ pub fn resolve(
             )?;
 
             // Phase 1 is photos-only: drop videos (a clip segment type lands
-            // in phase 2). Filter before sampling so the spread is over the
+            // in phase 2).
-            // photos that will actually appear.
             let items: Vec<memories::MemoryItem> = items
                 .into_iter()
                 .filter(|it| is_image_file(Path::new(&it.path)))
                 .collect();
 
-            let cap = (*max_segments).clamp(1, HARD_MAX_SEGMENTS);
+            // Years are derived from the whole span (what the reel represents),
-            let items = sample_evenly(&items, cap);
+            // before the budget narrows it down to beats.
-
             let years = distinct_years(&items, client_tz);
             let meta = ReelMeta { span: *span, years };
 
-            let planned = items
+            // The budget caps the number of narrated beats (≈ reel length);
-                .into_iter()
+            // each beat then bursts through several photos, so the reel covers
-                .map(|it| PlannedSegment {
+            // the span's moments without running minutes long.
-                    media: SegmentMedia::Photo {
+            let n_beats = budget_segments(*max_segments);
-                        rel_path: it.path,
+            let beats = form_beats(&items, n_beats, MAX_BURST_PHOTOS);
-                        library_id: it.library_id,
+            Ok((beats, meta))
-                    },
-                    date: it.created,
-                    insight_title: None,
-                    insight_summary: None,
-                })
-                .collect();
-            Ok((planned, meta))
         }
     }
 }
@@ -155,24 +277,24 @@ fn distinct_years(items: &[memories::MemoryItem], tz: Option<FixedOffset>) -> Ve
     years
 }
 
-/// Background pass: fill each segment's cached insight (title + summary) where
+/// Background pass: fill each beat's cached insight (title + summary) from its
-/// one exists. Best-effort — a missing or errored lookup leaves the fields
+/// lead photo, where one exists. Best-effort — a missing or errored lookup
-/// `None` and the scripter narrates from the date alone.
+/// leaves the fields `None` and the scripter narrates from the date alone.
 pub fn enrich(
     insight_dao: &Mutex<Box<dyn InsightDao>>,
     span_context: &opentelemetry::Context,
-    planned: &mut [PlannedSegment],
+    beats: &mut [PlannedBeat],
 ) {
     let Ok(mut dao) = insight_dao.lock() else {
         return;
     };
-    for seg in planned.iter_mut() {
+    for beat in beats.iter_mut() {
-        let rel_path = match &seg.media {
+        let Some(SegmentMedia::Photo { rel_path, .. }) = beat.photos.first() else {
-            SegmentMedia::Photo { rel_path, .. } => rel_path,
+            continue;
         };
         if let Ok(Some(insight)) = dao.get_insight(span_context, rel_path) {
-            seg.insight_title = Some(insight.title);
+            beat.insight_title = Some(insight.title);
-            seg.insight_summary = Some(insight.summary);
+            beat.insight_summary = Some(insight.summary);
         }
     }
 }
@@ -249,4 +371,78 @@ mod tests {
         ];
         assert_eq!(distinct_years(&items, None), vec![2019, 2021]);
     }
+
+    // Build an item at a given unix timestamp (seconds).
+    fn item_at(ts: i64, name: &str) -> memories::MemoryItem {
+        memories::MemoryItem {
+            path: format!("{name}.jpg"),
+            created: Some(ts),
+            modified: None,
+            library_id: 1,
+        }
+    }
+
+    #[test]
+    fn budget_segments_caps_to_duration_target() {
+        // 90s / 5s ≈ 18, bounded by the request max and hard cap.
+        assert_eq!(budget_segments(40), 18);
+        assert_eq!(budget_segments(5), 5); // request asked for fewer
+        assert_eq!(budget_segments(1000), 18); // hard cap / budget wins
+    }
+
+    #[test]
+    fn cluster_by_gap_splits_on_large_jumps() {
+        // Two photos minutes apart, then one a day later → two events.
+        let items = vec![
+            item_at(1_000_000, "a"),
+            item_at(1_000_300, "b"), // +5 min → same event
+            item_at(1_100_000, "c"), // +~27h → new event
+        ];
+        let clusters = cluster_by_gap(&items, EVENT_GAP_SECONDS);
+        assert_eq!(clusters.len(), 2);
+        assert_eq!(clusters[0].len(), 2);
+        assert_eq!(clusters[1].len(), 1);
+    }
+
+    #[test]
+    fn form_beats_one_beat_per_event_when_they_fit() {
+        // Three well-separated events, budget of 10 → three beats, each holding
+        // all of its (few) photos.
+        let items = vec![
+            item_at(0, "a"),
+            item_at(50, "b"), // same event as a
+            item_at(1_000_000, "c"),
+            item_at(2_000_000, "d"),
+        ];
+        let beats = form_beats(&items, 10, MAX_BURST_PHOTOS);
+        assert_eq!(beats.len(), 3);
+        assert_eq!(beats[0].photos.len(), 2); // burst of the first event
+        assert_eq!(beats[1].photos.len(), 1);
+        assert_eq!(beats[2].photos.len(), 1);
+    }
+
+    #[test]
+    fn form_beats_merges_events_when_over_budget() {
+        // Six distinct events but only two beats → adjacent events fold in, and
+        // every event's photos still appear (capped by the burst max).
+        let items: Vec<memories::MemoryItem> = (0..6)
+            .map(|i| item_at(i as i64 * 1_000_000, &format!("e{i}")))
+            .collect();
+        let beats = form_beats(&items, 2, MAX_BURST_PHOTOS);
+        assert_eq!(beats.len(), 2);
+        let shown: usize = beats.iter().map(|b| b.photos.len()).sum();
+        assert_eq!(shown, 6); // all six moments still shown across two beats
+    }
+
+    #[test]
+    fn form_beats_caps_burst_to_max() {
+        // One dense event of 30 photos, generous budget → a single beat that
+        // bursts at most MAX_BURST_PHOTOS, not all 30.
+        let items: Vec<memories::MemoryItem> = (0..30)
+            .map(|i| item_at(i as i64, &format!("p{i}")))
+            .collect();
+        let beats = form_beats(&items, 18, MAX_BURST_PHOTOS);
+        assert_eq!(beats.len(), 1);
+        assert_eq!(beats[0].photos.len(), MAX_BURST_PHOTOS);
+    }
 }