Recorded per-test traces, phase 1: results on TestPrune's own suite

Measured 2026-09-26 with FsHotWatch CLI 0.14.0-alpha.69, tests.traces = {"record": "full-runs", "fingerprintInputs": ["global.json", "Directory.Packages.props"]}, macOS arm64. The host was also running another repository's release gate at the time, so CPU numbers are noisy.

Runs

Run	Tree	Result	Recorded
confirm 1	both projects traced	exit 1: 10 TestPrune.Trace.Tests failures (see below)	b1a6f30e… both projects
confirm 2	TestPrune.Trace.Tests "traces": false	exit 0	2bdf1176… TestPrune.Tests
confirm 3	unchanged	exit 0: run-2 verdict reused, no tests ran	nothing new

Runs 1 and 2 recorded identical TestPrune.Tests traces (same counters, 848 complete). trace_runs.status is recorded with no reason for every run: preparation refused nothing.

TestPrune.Trace.Tests is not traced

Traced, the suite's outcome changes: 10 tests that pass untraced fail. The shadow bin replaces the app's own TestPrune.Trace.Recorder build with the weaver's bundled one, so the suite that tests the recorder runs against a different, live recorder:

Failure	Cause
ScaffoldTests stubs inert, DumpRoundTripTests counters	the process recorder is live
ShadowBinTests only repository assemblies are woven	the shadow bin ships the recorder without its PDB
InputCaptureTests / LaunchTests child-trace tests	the bundled recorder predates the child-scope fix
5 EndToEndTests (recorder-no-output)	the bundled shim gives the woven fixture's process the parent's scope, so the fixture run has no main dump

The run also overflowed (10,370 overflow hits, 259 of 280 tests recorder-overflow), because the suite's own tests swap the process recorder for small test recorders. The project is marked "traces": false. Phase 1 never changes a verdict, so the opt-out is required rather than a choice.

Bars (TestPrune.Tests, run 2bdf1176…)

Bar	Pass when	Measured	Verdict
census	traced ≥ 0.99; ambient < 0.1 % or listed	traced 1024/1024 (1.0000); ambient 0/14,771,003	pass
audit	ExtraTotal = 0; MissingUser explained	sampled 11, extra 15 (one test, see below)	fail: tool defect
overhead	CPU ratio ≤ 1.15	1.101 (untraced median 83.0 s, traced 91.4 s); traced max RSS 1,910.8 MiB	pass
file-census	ratio ≤ 0.05	0.048 (reads-repo 21, fail-outside 20, fail-in-repo 0)	pass
coverage parity	identical cobertura with traces removed	not measured: confirm here collects no coverage	open
PDB / JIT	prepare refused nothing	no refused run	pass

Incomplete traces (99 of 1024 tests; 848 of 947 stored traces complete)

Reason	Tests	Cause
unmapped-code <StartupCode$TestPrune-Falco> closures (no-document)	72	7 closure methods in FalcoRouteAnalysis / RouteStore resolve to no PDB document; joiner defect
unmapped-code F# exception types (type-not-indexed)	13	SymbolTraversalBudgetExceeded, SymbolTraversalDepthExceeded, ProjectAnalysisFailed, ProjectOptionsEmpty are not in the symbol index
child-process-untraced dotnet, sleep, sh, chmod	14	tests that start non-woven processes; expected by design

Audit

The one test with extras is the theory row rich types and patterns without anon records produce no diagnostics(body: "let f (p: int * string) = p"). Its 15 "extra" ids are everything it executes, its own method included, and the audit reports it as "not recorded when run alone". The isolated --filter-display-name run did not select the row, so the comparison was against nothing. It is an audit defect, not cross-test leakage. The missing entries are static constructors, closures and union cases that an earlier test already ran in the parallel run. The 8 missing user entries in SymbolOccurrenceStoreTests are AstAnalyzer record constructors and union-case getters. They run once per process, so in the parallel run an earlier test ran them first.

Defects found

Owner	Defect
TestPrune.Trace (released recorder)	the Process.Start shim gives a child with its own dump directory the parent's scope, and Launch.run passes inherited TESTPRUNE_TRACE_* to the child. Fixed in this change; FsHotWatch must bundle the next recorder
TestPrune.Trace	the shadow bin silently substitutes its bundled recorder for an app's own build of the same version
TestPrune.Trace	joiner: closures in <StartupCode$…> classes have no document
TestPrune.Core / joiner	F# exception declarations are not indexed, so their probes are unmapped
TestPrune.Trace	audit: a sampled test the isolated run does not select is counted as extras
TestPrune.Trace	census reports a project's last recorded run after the project opted out

Second run (2026-09-27)

Measured with FsHotWatch CLI 0.14.0-alpha.71, which bundles TestPrune.Core 13.2.2 (index schema 18: F# exception types are indexed) and TestPrune.Trace 0.2.0. The "traces": false opt-out on TestPrune.Trace.Tests is removed. Same tests.traces, macOS arm64. The host was loaded again (load average 17–39), so CPU numbers are noisy.

Runs

Run	Result	Recorded
confirm (fresh workspace, test-impact DB recreated at schema 18)	exit 0, green full-suite run: TestPrune.Tests 1033 passed + 1 skipped, TestPrune.Trace.Tests 303 passed	8af7b8f8…

trace_runs for 8af7b8f8…:

Project	Status	Reason
TestPrune.Tests	recorded	none
TestPrune.Trace.Tests	refused	the app ships its own TestPrune.Trace.Recorder build; tracing would replace it (its deps.json lists the recorder as a project)

The shadow bin now refuses TestPrune.Trace.Tests by itself. The suite runs untraced and stays green, with no config opt-out. The measurement verbs refuse it with the same reason, so the bars below cover TestPrune.Tests only.

Bars (TestPrune.Tests)

Bar	Pass when	First run (alpha.69, 2bdf1176…)	Second run (alpha.71, 8af7b8f8…)	Verdict
census: traced	≥ 0.99	1024/1024 (1.0000)	1030/1030 (1.0000)	pass
census: ambient	< 0.001, or listed	0/14,771,003	0/14,851,732	pass
audit	ExtraTotal = 0; MissingUser explained	sampled 11, extra 15 (display-filter defect)	sampled 11, extra 0; missing only cctor/gen	pass
overhead	CPU ratio ≤ 1.15	1.101 (83.0 s → 91.4 s); max RSS 1,910.8 MiB	1.068 (67.4 s → 72.0 s); max RSS 1,781.3 MiB; see below	pass
file-census	ratio ≤ 0.05	0.048 (reads-repo 21, fail-outside 20, fail-in-repo 0)	0.048 (reads-repo 21, fail-outside 20, fail-in-repo 0)	pass
coverage parity	identical cobertura with traces removed	not measured	lines identical, branches not; see "Coverage parity" below	fail
PDB / JIT	prepare refused nothing on a traced project	no refused run	0 invalid of 5,774 verified methods; the one refusal is the intended own-recorder refusal	pass

The theory row that produced the first run's 15 false extras (rich types and patterns without anon records produce no diagnostics(body: "let f (p: int * string) = p")) was sampled again and is now selected alone: extra 0, missing none. No missing user entries remain.

Incomplete traces

Reason	First run (tests)	Second run (tests)
unmapped-code <StartupCode$…> closures (no-document)	72	0
unmapped-code F# exception types (type-not-indexed)	13	0
child-process-untraced (dotnet, sleep, sh, chmod)	14	14 (dotnet 6, sleep 5, sh 2, chmod 1)
total incomplete / stored traces	99 / 947 (848 complete)	14 / 953 (939 complete)

unmappedIds is 0. The only incomplete traces left are tests that start non-woven processes, which is expected by design.

Overhead: an unexplained traced failure

The first overhead --reps 3 exited 1. The ratio was 1.049, but traced samples 2 and 3 exited 2 (untraced runs all exited 0). The verb discards the launched app's output, so the failing tests are unknown. The failure did not come back in 8 more traced launches: 2 back-to-back runs sharing one dump directory, 3 runs interleaved with untraced ones as the verb does, and a second overhead run (all 6 samples exit 0; its numbers are in the table). The fshw daemon was idle during the failing run, so no rebuild overlapped it. Host load average was 25–39 at the time.

Owner	Defect
TestPrune.Trace	overhead (via Launch.run) discards a launch's output, so a sample with a nonzero exit cannot be diagnosed. It should keep the output of a failing launch

What still blocks the phase-1 bar

Item	State
coverage parity	measured below: line coverage is identical, branch coverage is not
3 consecutive traced full runs (D6 census bar)	this run recorded one confirm
intermittent traced exit 2 under load	2 of 12 traced launches, not reproduced; open until a failing launch's output is kept

Follow-up (2026-09-27): overhead reruns and coverage parity

overhead and audit now print a nonzero-exit launch's output (TestPrune.Trace Unreleased), which fixes the defect above.

Overhead reruns with output capture

Run	Load average	Ratio	Untraced median	Traced median	Peak RSS	Exits
1	24–28	1.022	66.2 s	67.7 s	1,831.7 MiB	all 0
2	24–31	1.027	69.1 s	71.0 s	1,802.5 MiB	all 0
3	22–27	1.043	65.6 s	68.4 s	1,840.9 MiB	all 0
4	19–25	1.030	67.0 s	69.0 s	1,777.0 MiB	all 0

None of the 12 traced launches exited nonzero, so the earlier exit 2 did not recur. Over the day, 2 of 23 traced measurement launches exited nonzero, both in the same overhead run. It stays open: the next failure will print its tests.

Coverage parity

The second run's "confirm collects no coverage" was wrong. fshw collects coverage by default (tests.projects[].coverage defaults to true). Each confirm writes coverage/<project>/coverage.baseline.cobertura.xml per project and a merged, line-only coverage/coverage.cobertura.xml. Two confirm runs on the same tree, traced (run ef14ad88…, TestPrune.Tests recorded, TestPrune.Trace.Tests refused) and with tests.traces removed, were compared per file on lines covered/valid and branches covered/valid:

Report	Files	Line counts	Branch counts	Recorder module
TestPrune.Tests (traced)	31 = 31	identical in every file (5,504 / 5,903)	differ in 8 files: traced 1,196 / 1,548, untraced 1,283 / 1,646	none
TestPrune.Trace.Tests (refused, so untraced both times)	32 = 32	identical	identical (696 / 712)	none
merged coverage/coverage.cobertura.xml	63 = 63	identical (8,594 / 8,994)	none reported	none

The traced report adds no branch and changes no line hit. It drops every branch point on 31 lines, 98 branch points in all: AstAnalyzer.fs (38), Orchestration.fs (28), SymbolDiff.fs (16), Database.fs (8), AuditSink.fs, EdgeEmission.fs, ImpactAnalysis.fs, Program.fs (2 each). Every such line is a union match (match change with, match r.Outcome with) or a union-case test (if sym.IsExtern then). Those are the sites where the site-probe pass instruments the union's tag read before the switch. The loss is deterministic: two traced runs wrote byte-identical reports.

Owner	Defect
TestPrune.Trace (weaver, site probes)	coverage of a woven assembly loses the branch points of union matches and union-case tests: 98 of 1,646 in TestPrune.Tests. Line coverage is unaffected. A consumer whose coverage ratchet reads per-project branch rates from a traced run would see them drop. Cause not yet confirmed; the tag-read probe before the switch is the likely one

Coverage parity with the weaver fix (2026-09-27)

confirm collects per-project cobertura by default. A traced and an untraced confirm on alpha.71 matched on lines but not on branches: the traced TestPrune.Tests report lost 98 of 1,646 branch points (1,196 / 1,548 against 1,283 / 1,646), all on lines with a union match, a field comparison or a type test.

Cause

MS CodeCoverage counts a conditional branch under a hidden sequence point only while no call lies between it and its line's visible point. F# puts a match's or comparison's test under a hidden point, and a site probe is a call inserted right there, so the line keeps its hits and loses its branch points. Observed on the IL:

Unwoven shape	Branch counted
nop [L]; ldarg [hidden]; ldfld; …; ble	yes
… [hidden]; callvirt get_IsCircle; brfalse (an if x.IsCase)	no, even untraced
the same ldfld shape with a probe call after the ldfld	no
a probe, then a hidden copy of the enclosing point where the probe resumes	no
a probe, then a visible copy of the line's point where the probe resumes	yes

Fix (TestPrune.Trace, Unreleased)

Where a probe precedes a conditional branch in a hidden chain that falls through from its line (before any call, transfer or jump target), the site-probe pass gives the instruction the probe resumes at a copy of the line's visible sequence point. Probes and ids are unchanged; only the woven PDB gains points. CoverageParityTests weaves the FxTests fixture and requires every FxLib line's hits and condition-coverage to match the untraced run.

Measurement

The pinned fshw (alpha.71) bundles the released weaver, so confirm cannot measure an unreleased weaver fix. TestPrune.Tests was therefore woven with this tree's weaver and run twice with fshw's default coverage arguments (--coverage --coverage-output-format cobertura): the shadow apphost traced, and bin/Debug untraced. Both runs: 1,037 tests, 0 failed.

Report (repository files)	Lines covered / valid	Branches covered / valid	Files differing
untraced	5,552 / 5,949	1,283 / 1,644	—
traced, released weaver (confirm, alpha.71)	identical	1,196 / 1,548 (98 points lost)	8
traced, fixed weaver	identical, hits identical	1,279 / 1,640 (4 points lost)	1
traced, weaver closing the copy's range (below)	identical, hits identical	1,283 / 1,644 (0 points lost)	0

The 4 remaining points were Orchestration.fs lines 679 and 741 (match selection with): the list test of | RunSubset [] -> sits under a later hidden point, after calls, and counts unwoven. Rewriting only the PDB of a woven fixture of that shape (FxLib.Logic.selected) showed MS CodeCoverage drops such a branch when the range of the visible point heading its chain holds a call. The copy covered the probes that follow the case test, not just the case test. A hidden point right after the copy's last branch keeps its range call-free and restores the list test's points; hidden points later in the chain do not. The site-probe pass now adds that point, and CoverageParityTests covers the shape. Measured as above, with the weave cache cleared: the cache key hashes the woven inputs, not the weaver, so two local weaver builds of one version share a weave. The traced report has no TestPrune.Trace.Recorder package when the recorder's PDB is absent (the released recorder). A recorder built in this tree has its PDB at its build path, and MS CodeCoverage then reports it.

Bar	Pass when	Measured	Verdict
coverage parity	identical per-file lines and branches; no recorder module	lines identical; branches 1,640 / 1,644 (2 lines differ); no recorder module with the released recorder	fail (4 points), until a confirm on a release bundling the fix

Third run (alpha.74, 2026-09-27)

FsHotWatch CLI 0.14.0-alpha.74, which bundles TestPrune.Trace 0.3.0: the site-probe branch-point fix and the output capture of a failing measurement launch. Same tests.traces, macOS arm64, fresh workspace off main. Several other agents shared the host, so each launch is listed with its load average.

Runs

Run	Load (start → end)	Result	TestPrune.Tests	TestPrune.Trace.Tests
confirm 1 (1d0448d5…)	31.2 → 40.6	exit 0	1036 passed + 1 skipped; recorded	307 passed; refused (own recorder)
confirm 2 (be4a9b83…)	23.6 → 23.2	exit 0	1036 passed + 1 skipped; recorded	307 passed; refused (own recorder)
confirm 3 (165841b2…)	21.6 → 37.3	exit 0	1036 passed + 1 skipped; recorded	307 passed; refused (own recorder)
confirm, tests.traces removed (21e976a7…)	43.7 → 29.7	exit 0	1036 passed + 1 skipped	307 passed

.fshw.json was restored byte-identical after the untraced run.

Recorded stats, TestPrune.Tests	Run 1	Run 2	Run 3
executed / traced / complete	1033 / 1033 / 942	1033 / 1033 / 942	1033 / 1033 / 942
test-scope hits	19,389,838	19,389,838	19,389,838
static-init / ambient / overflow hits	410 / 0 / 0	410 / 0 / 0	410 / 0 / 0
unmapped ids, rejected dumps	0, none	0, none	0, none
recorder CPU	67.7 s	61.9 s	58.1 s

Bars (TestPrune.Tests; TestPrune.Trace.Tests is refused by every verb with the own-recorder reason, exit 2)

Bar	Pass when	Measured	Load	Verdict
census: traced	≥ 0.99 on 3 consecutive full runs	1033/1033 (1.0000) on runs 1, 2 and 3	see Runs	pass
census: ambient	< 0.001, or listed	0 / 19,390,250 on each run	see Runs	pass
audit	ExtraTotal = 0; MissingUser explained	sampled 11, extra 0; missing only cctor (and 1 gen); no missing user	29.1 → 29.0	pass
overhead	CPU ratio ≤ 1.15	1.066 (untraced median 75.8 s, traced 80.8 s); peak RSS 1,872.9 MiB; all 6 samples exit 0	19.7 → 21.2	pass
file-census	ratio ≤ 0.05	0.048 (reads-repo 21, fail-outside 20, fail-in-repo 0)	29.0 → 19.7	pass
coverage parity	identical per-file lines and branches; no recorder module	lines and every line's hit count identical in all 63 files; branches differ in 1 file (4 points, below); no recorder module in the traced report	43.7 → 29.7	fail (4 points, known)
PDB / JIT	prepare refused nothing on a traced project	trace_runs.status recorded, no rejected dump, on all 3 runs; the only refusal is the intended own-recorder one	—	pass

Coverage parity (run 3 against the untraced run)

Report	Files	Lines covered / valid	Branches covered / valid	Differing files
TestPrune.Tests traced	31 = 31	5,552 / 5,949 =	1,279 / 1,640 against 1,283 / 1,644	1
TestPrune.Trace.Tests (refused, untraced both times)	32 = 32	3,160 / 3,161 =	718 / 734 =	0
merged coverage/coverage.cobertura.xml	63 = 63	8,712 / 9,110 =	none reported	0

File	Line	Traced	Untraced	Explanation
src/TestPrune/Orchestration.fs	679	hits 1, branches 2/2	hits 1, branches 4/4	the recorded gap after the weaver fix (match selection with, a branch under a later hidden point); being worked separately
src/TestPrune/Orchestration.fs	741	hits 1, branches 2/2	hits 1, branches 4/4	same

The alpha.71 traced report lost 98 branch points in 8 files. alpha.74 loses the 4 above and nothing else. The three traced runs wrote identical per-file reports. TestPrune.Trace.Tests reports a TestPrune.Trace.Recorder package traced and untraced alike: that suite tests the recorder, and it is refused, so the package is its own code under test, not a woven-in module.

Other observations

Observation	Detail
no traced launch exited nonzero	3 traced confirms, 3 traced overhead samples, the audit's parallel and 11 isolated launches: all exit 0. The intermittent exit 2 did not recur at load 20–40
confirm on an unchanged tree cannot be forced to re-run	confirm, invalidate + confirm and --no-cache confirm --run-once all replay the previous verdict (.fshw/verdict.json and the daemon's receipt). Runs 2 and 3 were forced by stopping the daemon and moving verdict.json aside
census of an older run loses its incomplete reasons	census --run 1d0448d5… after run 2 prints no incomplete line; the same tests' traces were re-recorded by run 2, which replaced run 1's reasons. The newest run prints child-process-untraced 14
audit's parallel run has 17 child-process-untraced tests, the census 14	audit: dotnet 9, sleep 5, sh 2, chmod 1; confirm run: 14 (second run's breakdown: dotnet 6, sleep 5, sh 2, chmod 1). The 3 extra dotnet children are in the audit's own launch only; not investigated

Fourth run (alpha.78, 2026-09-29)

FsHotWatch CLI 0.14.0-alpha.78, which bundles TestPrune.Core 13.3.0 and TestPrune.Trace 0.4.1. That includes 0.3.1's fix for branch points under a match's later hidden sequence points and 0.4.0's static-init inheritance. Same tests.traces, macOS arm64 on AC power, a fresh workspace off main with only the pin bump. Each traced full run is dotnet fshw confirm --fresh. The measurement verbs ran from this tree's src/TestPrune.Trace.Cli, whose TestPrune.Trace is the released 0.4.1. CPU busy is the mean over two 1 s top samples just before each launch.

Runs

Run	Start (UTC)	Load at start (1/5/15 min)	CPU busy	Result	TestPrune.Tests	TestPrune.Trace.Tests
confirm 1 (4d10d1f7…)	12:24:14	9.7 / 13.0 / 38.3	93 %	exit 0	1041: 0 failed, 1 skipped; recorded	328: 0 failed; refused (own recorder)
confirm 2 (9240123e…)	12:30:27	29.0 / 38.1 / 43.5	72 %	exit 0	1041: 0 failed, 1 skipped; recorded	328: 0 failed; refused (own recorder)
confirm 3 (be272951…)	12:48:51	9.7 / 21.4 / 30.7	35 %	exit 0	1041: 0 failed, 1 skipped; recorded	328: 0 failed; refused (own recorder)
untraced full run (6cd21e06…), tests.traces removed	12:49:48	14.5 / 21.2 / 30.1	21 %	exit 0	1041: 0 failed, 1 skipped	328: 0 failed

All three traced runs have the same input tree (sha256:8e2e8784…). .fshw.json was restored byte-identical after the untraced run.

Recorded stats, TestPrune.Tests	Run 1	Run 2	Run 3
executed / traced / complete	1037 / 1037 / 946	1037 / 1037 / 946	1037 / 1037 / 946
test-scope hits	19,785,966	19,785,966	19,785,966
static-init / ambient / overflow hits	770 / 0 / 0	770 / 0 / 0	770 / 0 / 0
unmapped ids, rejected dumps	0, none	0, none	0, none
recorder CPU	70.2 s	67.7 s	70.1 s

Against alpha.74 there are 946 complete traces instead of 942, and 770 static-init hits instead of 410. Both changes come from 0.4.0, which records each type initializer in its own scope and links it to the tests that use the type.

Bars (TestPrune.Tests; every verb refuses TestPrune.Trace.Tests with the own-recorder reason, exit 2)

Bar	Pass when	Measured	Start (UTC), load, CPU busy	Verdict
census: traced	≥ 0.99 on 3 consecutive full runs	1037/1037 (1.0000) on runs 1, 2 and 3	see Runs	pass
census: ambient	< 0.001, or listed	0 / 19,786,736 on each run	see Runs	pass
audit	ExtraTotal = 0; MissingUser explained	sampled 11, extra 0; missing only cctor; no missing user	12:50:48, 17.2, 17 %	pass
overhead	CPU ratio ≤ 1.15	1.067 (untraced median 78.1 s, range 78.0–78.2; traced 83.4 s, range 81.8–84.5); peak RSS 1,856.5 MiB; all 6 samples exit 0	12:51:40, 9.9, 22 %	pass
file-census	ratio ≤ 0.05	0.000 (fail-outside 20, reach-outside 1, reads-repo 21, fail-in-repo 0)	12:54:42, 6.2, 30 %	pass
coverage parity	identical per-file lines and branches; no recorder module	identical in every file, line hit counts included (below)	see Runs	pass
PDB / JIT	prepare refused nothing on a traced project	trace_runs.status recorded, no rejected dump, on all 3 runs; the only refusal is the intended own-recorder one	—	pass

The file-census ratio fell from 0.048 to 0. The one test that used to account for the difference reads a repository file at an absolute path. 0.4.0's file-census now counts that read, made from outside the repository, as a dependency on the repository (reach-outside 1).

Coverage parity (traced run 3 against the untraced run)

Report	Files	Lines covered / valid	Branches covered / valid	Differing files
TestPrune.Tests traced	31 = 31	5,561 / 5,957 =	2,160 / 2,690 =	0
TestPrune.Trace.Tests (refused, untraced both times)	33 = 33	3,250 / 3,251 =	1,051 / 1,067 =	0
merged coverage/coverage.cobertura.xml	64 = 64	8,811 / 9,208 =	none reported	0

Every line's hit count and condition coverage is identical as well. The traced TestPrune.Tests report has no TestPrune.Trace.Recorder package. TestPrune.Trace.Tests reports one traced and untraced alike, because that suite's code under test is the recorder. The totals are not comparable with earlier rounds (the tree and the pinned SDK band both changed since then); parity is measured within this round.

Observations

Observation	Detail
audit and census count child-process-untraced differently	explained: the audit's parallel run counts 9 dotnet tests and the census 6. The census stores one trace per test method (class + method), while the audit counts each xUnit test instance. NuGetPublicationBarrierTests starts dotnet from 5 facts and from one theory with 4 InlineData rows: 5 + 1 = 6 methods, 5 + 4 = 9 instances. The other kinds (sleep 5, sh 2, chmod 1) have no theories, so they agree. Nothing to fix
the untraced full run was launched by the daemon, not by confirm	removing tests.traces changed .fshw.json, so the daemon stopped and a new daemon ran the full suite (6cd21e06…, coverage written 12:50:41). confirm --fresh, started four seconds after the edit, graded that run and printed "no test needed: nothing changed since run 6cd21e06…, which verified this tree". The run was real and complete; only the wording reads like a skip
census of an older run loses its incomplete reasons	unchanged from the third run: re-recording a test replaces its reasons
no traced launch exited nonzero	3 traced confirms, 3 traced overhead samples, and the audit's parallel run and 11 isolated launches

Phase-1 bar on TestPrune's own suite

Every D6 bar holds for TestPrune.Tests on a released toolchain. TestPrune.Trace.Tests is intentionally not traced (ADR 0008). No unexplained item remains, so no defect was filed.