Fix Interstitial snap-out at live edge and BUFFER_APPEND_NO_PROGRESS false positives

[robwalch]

This PR will...

• Fix Interstitial start dates snapping to the last segment when they first appear at a distance from the live edge
• Fix BUFFER_APPEND_NO_PROGRESS (Mark fragments as gaps after repeated appends without buffered range growth #7941) false positives in asset buffering
• Ensure Interstitials are played at 1x over live primary

Why is this Pull Request needed?

These issues interfere with append-in-place buffering across scheduled mid-rolls at a low-latency live edge.

Are there any points in the code the reviewer needs to double check?

[DRAFT] TODO: Improve primary buffer resumption at low-latency edge

Resolves issues:

Fixes #7978

[rudemateo]

This early return also disables the "count parsed fragments that produced no append operations" protection from #7941

I think that checking buffering stats should fix the same false positive (appends were queued but displaced before completing by transfer/end-of-stream), and the original tests should then pass unmodified:

-    if (!progress) {
-      // Tracking miss
-      return;
-    }
-    const cycle = progress.stats === frag.stats ? progress : undefined;
+    const cycle = progress?.stats === frag.stats ? progress : undefined;
     if (cycle?.errored) {
       // Counted by the append-error path
       return;
     }
+    const fragBuffering = frag.stats.buffering;
+    if (fragBuffering.start > 0 && fragBuffering.first === 0) {
+      // Queued appends never completed (displaced by end-of-stream or transfer)
+      return;
+    }

(A per-SourceBuffer in-flight counter would potentially be even more precise, but I didn't want to suggest introducing new stuff if you think the buffering stats we already have are enough)

[rudemateo]

Tested this branch against the stream in #7978. The resumption fix works: parts at the item start load as soon as they're published, and no more catchup rate creep during breaks.

One issue that reproduces now that primary buffers past the break: the asset ends ~33ms short of the resume point, and the gap-controller nudge out of that hole can land just before the boundary of the item that ended a few ms earlier. The position handler treats it as a backward seek and restarts the interstitial. With X-CONTENT-MAY-VARY=YES one session got an empty asset list back and the break collapsed:

138.649 [interstitials]: Stalled at 7.8899 of 8 in HlsAssetPlayer: ["synthmid-...259200-2" 769.60-777.60]
138.653 [interstitials]: INTERSTITIAL_ENDED ["synthmid-...259200" 753.60-777.60]
138.663 [gap-controller]: Nudging 'currentTime' from 777.489901 to 777.589901
138.672 [interstitials]: INTERSTITIAL_STARTED ["synthmid-...259200": 753.60-777.60]
138.755 [interstitials]: Loaded asset-list with duration: 0 (was: 24)

Ignoring regressions within a buffer hole of the item start (onSeeking in interstitials-controller, line 953) fixes it here without affecting seek-back replay:

     if (
-      (backwardSeek && currentTime < playingItem.start) ||
+      (backwardSeek &&
+        playingItem.start - currentTime > this.hls.config.maxBufferHole) ||
       currentTime >= playingItem.end
     ) {

The narrow edge-gate suggestion comment I left in interstitials-controller also prevents the nudge on this stream, so these should be complementary.

[rudemateo]

Would <= config.maxBufferHole make sense here?

[rudemateo]

is this needed? it looks like this.config was removed from the constructor

[rudemateo]

Regarding the resumption TODO: I tested this branch with just the gate narrowed so primary resumes as soon as the playlist reaches the resume point

-      if (details?.live && bufferPos + 0.5 >= details.edge) {
+      if (
+        details?.live &&
+        bufferPos - details.edge > ALIGNED_END_THRESHOLD_SECONDS
+      ) {

combined with Line 1751 changed back to bufferedPos < details.edge.

Breaks in the #7978 stream are segment-aligned, so it looks like the +0.5 costs a full reload cycle at every cue-in. With the narrow gate the primary append merges across the boundary before the playhead gets there, and the end-of-asset stall/nudge shouldn't happen.