Two claims in E9 went further than the evidence, in an entry whose whole subject
is a figure that was quoted past its own.
"All 16 sat on that one line" was consistent with what was measured, not
established by it: `MediaProbe:321` accounts for a 16-branch difference and the
denominators now agree at 1338, but the other lines were never enumerated in both
reports, so a line that lost coverage elsewhere would have been invisible to the
check that was run. Says that now.
And `matroskaOrWebm` is not "the only string-literal `when` in that file" --
`shortName` at :429 is a second one. The parenthetical was never checked; it is
gone rather than repaired, since the mechanism was not what the measurement
established anyway.
The previous commit message carries the stronger wording. It is left alone rather
than rewritten, because that would need a force push.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
E8 classified the 32 lines neither suite executes and stopped there. It never
asked which *arms* neither suite takes on lines both suites run, and that tier
is where what is left actually lives.
Rebuilt the union rather than reusing E8's artifact: a fresh jacocoTestReport
merged with E8's own API 34 .ec against one set of current class files. Union is
99.0% line, 90.1% branch. Controls confirm the device half applied -- 32 lines in
FFmpegEngine, 24 in Media3Engine, 15 in ConcatEngine, 9 in MainActivity that the
JVM suite never reaches.
Three things came out of it that are worth more than the count:
- **E8's denominator artefact is retired.** It warned the union's branch
denominator ran 16 ahead "entirely inside MediaProbe" and told readers not to
quote a MediaProbe branch figure raw. Rebuilt, both denominators are 1338 and
MediaProbe:321 reads mb=0 cb=4. All 16 sat on that one line, so it was a
property of how the report was built and never of the code.
- **CLAUDE.md's "18" is not stale.** It reproduces exactly under `mi == 0 && mb > 0`
(19 branches on 18 lines) and not under `ci > 0`, which admits partially-executed
signature lines and gives 139. Different metrics, not drift -- stated in E9 so
the next read does not "correct" a figure that is right.
- **Tier 1 is 22, down from 32.** #252 closed the ten getForegroundInfo lines and
added none. Two classes carry a one-line blind spot because #252 changed their
bytecode and JaCoCo rightly rejected E8's .ec for them; the bound is measured,
not assumed, and the line is not a gap.
Of the 22 branch sites, 12 are decided here -- compiler codegen, an F4/F6/F10
exemption already on record, or a thread race -- and recorded in E9 rather than as
new F-entries, since coverage-read-findings.md is being rewritten by #261. The
other 10 are filed as #262-#266, each naming the mutation that must go red or
saying that the read is the ticket.
FFmpegCommandBuilder:183's missed arm is VORBIS, which #261 is closing as it
lands, so the set is 21 the day it merges. E9 says to re-derive rather than edit
that sentence.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
JMR-devandClaude Opus 5