improve: retune scrub supersession span default

Co-authored-by: Richie McIlroy <richiemcilroy@users.noreply.github.com>

Author: cursoragent

crates/editor/PLAYBACK-BENCHMARKS.md

@@ -95,7 +95,7 @@ cargo run -p cap-editor --example scrub-benchmark -- --video /path/to/video.mp4
 # Runtime tuning for FFmpeg scrub supersession heuristic
 CAP_FFMPEG_SCRUB_SUPERSEDE_MIN_PIXELS=3686400 \
 CAP_FFMPEG_SCRUB_SUPERSEDE_MIN_REQUESTS=8 \
-CAP_FFMPEG_SCRUB_SUPERSEDE_MIN_SPAN_FRAMES=25 \
+CAP_FFMPEG_SCRUB_SUPERSEDE_MIN_SPAN_FRAMES=20 \
 cargo run -p cap-editor --example scrub-benchmark -- --video /path/to/video.mp4
 
 # Export per-run and aggregate scrub metrics to CSV
@@ -178,6 +178,39 @@ cargo run -p cap-recording --example playback-test-runner -- full
   - `cargo run -p cap-editor --example playback-startup-report -- --log crates/editor/PLAYBACK-BENCHMARKS.md --run-id sample-run`
   - Completed successfully with filtered metric output path active.
 
+### Benchmark Run: 2026-02-14 00:00:00 UTC (supersession span retune to 20)
+
+**Environment:** Linux runner with synthetic 4k60 and 1080p60 MP4 assets  
+**Commands:** `scrub-benchmark --runs 3`, `playback-benchmark --seek-iterations 10`, `decode-benchmark --seek-iterations 10`  
+**Change under test:** default `CAP_FFMPEG_SCRUB_SUPERSEDE_MIN_SPAN_FRAMES` fallback changed from `25` to `20`
+
+#### 4k scrub span sweep before promoting new default
+- Command family:
+  - `CAP_FFMPEG_SCRUB_SUPERSEDE_MIN_SPAN_FRAMES={15,20,25,30} scrub-benchmark --runs 3`
+- Median last-request latency by span:
+  - **15**: avg **836.94ms**, p95 **1740.74ms**
+  - **20**: avg **814.93ms**, p95 **1743.49ms**
+  - **25**: avg **819.11ms**, p95 **1762.74ms**
+  - **30**: avg **923.18ms**, p95 **1947.86ms**
+- Decision: promote span **20** as new default based on best median average and lower tail than span 25/30.
+
+#### Scrub Benchmark — default after retune (span=20)
+- 4k60 (`/tmp/cap-bench-4k60.mp4`, runs=3):
+  - Median all-request: avg **832.56ms**, p95 **1732.40ms**, p99 **1732.40ms**, max **1732.41ms**
+  - Median last-request: avg **836.61ms**, p95 **1732.40ms**, p99 **1732.40ms**, max **1732.40ms**
+- 1080p60 (`/tmp/cap-bench-1080p60.mp4`, runs=3):
+  - Median all-request: avg **222.58ms**, p95 **446.05ms**, p99 **472.21ms**, max **472.21ms**
+  - Median last-request: avg **326.36ms**, p95 **472.21ms**, p99 **472.21ms**, max **472.21ms**
+
+#### Regression checks after default retune
+- Playback throughput:
+  - 1080p60: **60.24 fps**, missed deadlines **0**, decode p95 **2.24ms**
+  - 4k60: **60.18 fps**, missed deadlines **2**, decode p95 **9.67ms**
+- Decode benchmark:
+  - 1080p random access avg **111.79ms**, p95 **337.65ms**
+  - 4k random access avg **509.26ms**, p95 **1451.87ms**
+- Duplicate burst handling remained stable (0 failures for burst sizes 4/8/16).
+
 ### Benchmark Run: 2026-02-14 00:00:00 UTC
 
 **Environment:** Linux runner with synthetic 1080p60 and 4k60 MP4 assets  

crates/editor/PLAYBACK-FINDINGS.md

@@ -831,6 +831,44 @@ The CPU RGBA→NV12 conversion was taking 15-25ms per frame for 3024x1964 resolu
 
 ---
 
+### Session 2026-02-14 (Supersession span threshold retune to 20)
+
+**Goal**: Re-evaluate supersession span threshold with CSV-backed multi-run sweeps and improve 4k scrub medians
+
+**What was done**:
+1. Ran a 4-way sweep over `CAP_FFMPEG_SCRUB_SUPERSEDE_MIN_SPAN_FRAMES={15,20,25,30}` with `scrub-benchmark --runs 3`.
+2. Compared median last-request latency and p95 tails from CSV outputs.
+3. Updated FFmpeg supersession default span fallback from `25` to `20`.
+4. Re-ran scrub, playback, and decode regression benchmarks after the default change.
+
+**Changes Made**:
+- `crates/rendering/src/decoder/ffmpeg.rs`
+  - changed default `min_span_frames` fallback from `25` to `20`
+- `crates/editor/PLAYBACK-BENCHMARKS.md`
+  - updated command examples to show span `20`
+  - added benchmark entry for the threshold sweep and post-change regression runs
+
+**Results**:
+- 4k sweep medians (last-request avg / p95):
+  - span 15: **836.94ms / 1740.74ms**
+  - span 20: **814.93ms / 1743.49ms**
+  - span 25: **819.11ms / 1762.74ms**
+  - span 30: **923.18ms / 1947.86ms**
+- Post-change default (span 20) validation:
+  - 4k scrub median last-request avg **836.61ms**, p95 **1732.40ms**
+  - playback throughput remains 60fps-class:
+    - 1080p: **60.24 fps**
+    - 4k: **60.18 fps**
+  - decode metrics remain in expected variance envelope:
+    - 1080p random avg **111.79ms**
+    - 4k random avg **509.26ms**
+
+**Decision**: keep defaults at `min_requests=8`, `min_span_frames=20`.
+
+**Stopping point**: supersession defaults now favor a slightly more aggressive span threshold while preserving 60fps throughput and stable decode behavior.
+
+---
+
 ### Session 2026-02-14 (Rejected superseded-burst cache-window reduction)
 
 **Goal**: Reduce superseded scrub decode work by shrinking decode cache window for superseded requests

crates/rendering/src/decoder/ffmpeg.rs

@@ -102,7 +102,7 @@ fn scrub_supersession_config() -> ScrubSupersessionConfig {
             .unwrap_or(8);
         let min_span_frames = parse_u32_env("CAP_FFMPEG_SCRUB_SUPERSEDE_MIN_SPAN_FRAMES")
             .filter(|value| *value > 0)
-            .unwrap_or(25);
+            .unwrap_or(20);
         let min_pixels = parse_u64_env("CAP_FFMPEG_SCRUB_SUPERSEDE_MIN_PIXELS")
             .filter(|value| *value > 0)
             .unwrap_or(3_686_400);