PrismCoreEngine

The default per-source prewarm budget, and its ceiling: half the coordinated reader's 4 MB retention, so the reader can take over everything that was fetched and still have room for its own fills.

Whether this build can re-encode non-streamable audio to EAC3.

Public because it changes what PrismCore can do, and a host that ships its own FFmpeg build is the one that decides the answer: stock MPVKit has the ac3/eac3 decoders only, so the bridge is unavailable until a fork configures --enable-encoder=eac3.

Probe url and decide which engine takes it — without starting anything.

Exposed separately from open so a host can log or gate on the decision, and so the routing rules are testable without standing up a server or a renderer.

Fetch the bytes a play of url will read first — the container header and, for a Matroska, the Cues at its tail with the last cluster before them — so a later open over the coordinated HTTP reader (coordinatedHTTP: true) starts from memory.

For a host that knows what is likely next: the following episode, the item under the cursor, a detail page. Safe to call speculatively and to ignore: the result is advisory, and a prewarm nobody plays is dropped by the store's budget or the first memory warning.

What makes it safe to use. Before the first prewarmed byte is delivered, the reader asks the origin once more (a one-byte range) and takes the blocks only if the strong ETag and the length are what they were. An origin that reports no strong ETag — including one that reports only Last-Modified — cannot be prewarmed at all; see SourcePrewarmOutcome.Status.validatorUnavailable.

What it costs the origin. Every request goes through the same per-origin admission as playback, at lower priority: never while another request to the origin is in flight, and not at all within a minute of a refusal. A 429 / 503 is recorded, not retried.

• Parameters:

  • httpHeaders: the headers the play will send. Part of the identity: a prewarm under one token is not offered to a play under another.
  • byteBudget: the most this prewarm may hold. Clamped to defaultPrewarmByteBudget, which is all a reader can retain.

Drop every prewarmed source — for a host that knows the user left the screen the prewarms were for, or that wants the memory back first.

Probe, decide, and start the chosen engine.

The probe is a real open — a network round trip on a server source — and starting either engine blocks on I/O, so this is async and must not be called from a context that can't wait.

• Parameters:

  • display: what the display can present. Only the remux path uses it (it gates the HDR and Dolby Vision claims in the master playlist); the software path renders into a layer and makes no claims.
  • segmentCacheBytes: remux path only — disk budget for produced segments.

Start the software pipeline over an already-probed source — the entry a host uses when it ran SourceProbe.open + decide itself (to log or gate on the decision) and wants the probe's open to count. The pipeline adopts the context; a ProbedSource already consumed elsewhere makes it open probed.url afresh instead.

Returned loaded and paused, like open(url:)'s software case.

A report for a source that never got a session — declined, routed to the software path, or simply not played — from its probe and the routing outcome:

let probed = try await SourceProbe.openDetached(url: url, httpHeaders: headers)
let outcome = Result { try PrismCoreEngine.decide(for: probed.info) }
let report = PrismCoreEngine.diagnosticReport(for: probed, decision: outcome)

options and runtime are absent: there was no session to have them.