C2 (#177): cut MediaProbe's two-probe merge into a seam, and ask which probe wins

probe() runs MediaExtractor and FFprobe independently and merges the two, and every rule
in that merge is a decision nothing held. The reason is structural rather than an
oversight: RemuxTest drives the whole thing on a device against committed fixtures, but
only ever with one probe answering and the other agreeing or also failing. Nothing on any
source set can arrange for a real extractor and a real FFprobe to *disagree*, so every
elvis in the merge was taken in one direction and never the other.

The seam is `internal fun merge(Extracted?, FFprobeInfo?): InputProbe`, pulled out of
probe() whole -- probe() now reads the two probes, merges, and keeps the log. FFprobeInfo
becomes internal alongside it; Extracted already was, with a KDoc giving this exact reason,
and FFprobeInfo simply never got the same treatment. Half a signature being private is
what made the function unnameable from a test.

Eleven tests, and the mutations that hold them:

  image beats a real video codec      demote the isImage arm below the video arm
  the extractor wins on codecs        flip the elvis to FFprobe-first
  duration is the larger reading      replace maxOf with extractor-first
  dimensions prefer the extractor     flip the width elvis
  no recognised stream is unreadable  narrow the guard to `extracted == null && info == null`

All five red, then restored. One mutation I tried first was *semantically equivalent* --
moving the image arm above the both-null arm changes nothing for any reachable input -- so
it stayed green and is recorded here rather than counted: a green mutation is only evidence
when the mutation is a real change.

The last row is the arm the ticket was filed for: parsed, and carrying no stream either
probe recognised, which is what a container holding only subtitles looks like. Its input
was already being constructed elsewhere in the suite -- MediaProbeTrackWalkTest calls
extractedFrom(emptyList()) and gets exactly it -- and had never been handed to the merge.

568 -> 579 JVM tests, 0 failures.
MediaProbe: 35 -> 24 missed lines, 72 -> 40 missed branches.
Line 2103/2348 -> 2173/2348; branch 1029/1340 -> 1091/1342.

The branch denominator moved by two, and it is the seam that moved it -- worth stating
separately from the numerator, because CLAUDE.md's coverage entry has a documented history
of explaining its own numbers wrongly.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-09-01 21:59:00 -05:00
co-authored by Claude Opus 5
parent eded47d666
commit 794cef7b34
2 changed files with 203 additions and 10 deletions
@@ -50,19 +50,42 @@ object MediaProbe {
)
fun probe(context: Context, uri: Uri): InputProbe {
val extracted = probeWithExtractor(context, uri)
val info = probeWithFFprobe(context, uri)
val merged = merge(probeWithExtractor(context, uri), probeWithFFprobe(context, uri))
if (merged.kind == InputKind.UNPARSEABLE) {
// Not a failure: an unparseable input is a strong signal that this job belongs on
// FFmpeg. Reporting an unknown codec makes the router say so.
Log.i(TAG, "Neither MediaExtractor nor FFprobe could read $uri; routing to FFmpeg.")
}
return merged
}
/**
* What the two probes together say about one input.
*
* A pure function, and `internal` for the same reason [extractedFrom] is: the precedence rules
* below are the answer to "which probe wins", and until this was pulled out of [probe] the only
* way to ask was to have a real `MediaExtractor` and a real FFprobe **disagree**, which nothing
* on any source set can arrange. `RemuxTest` drives this on a device against committed
* fixtures, but only ever with one probe answering and the other agreeing or also failing --
* so every elvis here was taken in one direction and never the other.
*
* The rules, each of which is a decision rather than an accident:
*
* - **The extractor wins on codecs.** It is the platform's own view of what it can decode,
* which is the thing the router is about to ask about. FFprobe's name for the same track can
* differ, and the copy planner keys off these strings.
* - **FFprobe alone reports the container.** `MediaExtractor` cannot, which is why [InputProbe]
* carries a nullable one and `CopyPlanner` treats null as "container unknown".
* - **Duration is the larger of the two**, not the first non-zero. Either probe can report zero
* for a file the other times correctly, and a zero duration makes the FFmpeg progress
* percentage undefined.
*/
internal fun merge(extracted: Extracted?, info: FFprobeInfo?): InputProbe {
val videoCodec = extracted?.videoCodec ?: info?.videoCodec
val audioCodec = extracted?.audioCodec ?: info?.audioCodec
val kind = classify(extracted, info)
if (kind == InputKind.UNPARSEABLE) {
// Not a failure: an unparseable input is a strong signal that this job belongs on
// FFmpeg. Reporting an unknown codec makes the router say so.
Log.i(TAG, "Neither MediaExtractor nor FFprobe could read $uri; routing to FFmpeg.")
return UNREADABLE
}
if (kind == InputKind.UNPARSEABLE) return UNREADABLE
return InputProbe(
videoCodec = videoCodec,
@@ -83,7 +106,7 @@ object MediaProbe {
* audio file and a corrupt file indistinguishable. The source-info card cannot describe either
* honestly until they are separate, and neither can the copy planner.
*/
private fun classify(extracted: Extracted?, info: FFprobeInfo?): InputKind = when {
internal fun classify(extracted: Extracted?, info: FFprobeInfo?): InputKind = when {
info?.isImage == true -> InputKind.IMAGE
extracted == null && info == null -> InputKind.UNPARSEABLE
(extracted?.videoCodec ?: info?.videoCodec) != null -> InputKind.VIDEO
@@ -162,7 +185,11 @@ object MediaProbe {
}
}
private class FFprobeInfo(
/**
* `internal` rather than `private` for the same reason [Extracted] is, and it should have been
* from the start: [merge] cannot be named from a test while half its signature is private.
*/
internal class FFprobeInfo(
val container: Container?,
val videoCodec: String?,
val audioCodec: String?,