Skip to content

Releases

Subpoena v0.1.0-alpha.6

Pre-release

Choose a tag to compare

@nickhenggeler nickhenggeler released this 10 Nov 00:30
bd3ebca

Tagged v0.1.0-alpha.5 and v0.1.0-alpha.6, on the same commit: the fixes and the new witnesses
landed interleaved, so no earlier commit holds the fixes alone. No Python release goes with these
tags. PyPI stays at 0.1.0a4 until a python/ tag is cut.

What is new:

  • Three witnesses: file writes, database commits (Postgres and MySQL, experimental) and
    delivery on an inbound socket.
  • A kill has to hold: it reruns on the whole suite, with and without the mutant, and a Go kill
    needs a failed test that links the mutated package.
  • Sinkhole can answer 200, and Marked tells one arrival from another.
  • Cross output prints upstream line numbers and counts markers against mutants.
  • Python verdicts: --test-helper declares test code, Mock.assert_* under pytest-mock is a
    kill, and production code run by --doctest-modules no longer counts as test code.

Stated limits:

  • Forged runner output can fool a plain kill. Only a verified witness mark under
    --require-witness holds against it (docs/architecture.md, "What forged output can do").

  • A witness mark does not hold against code that reads the key and signs its own mark. Do not rely
    on one against a suite written to cheat.

  • A Pass from the database witness is experimental, and its known gaps are listed in its entry below.

  • A poll inside an exec waits for the arguments. ProcWitness could land after an exec set the
    new task name and before the arguments existed, and it kept that name as an unidentified event for
    good. Under an Allowed list that went red with nothing wrong, in about 2 of 100 batches of the
    gateway suite under load. Such a poll now re-reads the process for up to 5ms, and takes the line if
    the name, binary and cgroup still match. If they do not, or the wait runs out, the name stays and
    the check goes red as before. A read that pairs the old name with the new line can still go red.

  • A run that cost a kill says which test failed in it. An unconfirmed kill showed the last 20
    lines of the red run and the mutant run's failing tests. In a long suite those lines are the slowest
    tests' passes, so the test that failed without the mutant could not be seen. Where that run has a
    failure to name, the output now starts with its failed tests and first failure. What counts as unconfirmed is unchanged.

  • A panic testify's suite recovered is a crash. It reports one through t.Errorf, with no
    line at column 0, so it read as a plain kill. test panicked: after a file and line, in a test
    that did not pass, now counts as a crash, as Ginkgo's [PANICKED] does. New cost: a failed test
    that logs those words reads crashed.

  • A recovered panic in a parallel test is a crash. Package testing's own frames can put the
    panic: line of one parallel test on another that passed, where it read as a log line and the
    test's plain failure as a kill, with nothing forged. A crash line on a passing test now counts if
    that test ran beside one that did not pass, read from each test's events from run, or cont if
    it paused, to its end, whatever slashes the subtests' names hold. That covers subtests started from
    goroutines, with three exceptions that stay plain kills and are stated in docs/architecture.md: one whose name goes on from a
    sibling's, which passes and ends while that sibling runs, one whose parent carries on and recovers the panic in its own
    body, and one that works and then calls t.Parallel. Sequential suites
    keep the log-line rule, unless a TestMain that calls m.Run again repeats their names. New cost: a passing test that logs such a line beside a real
    failure reads crashed, and the run fails.

  • A quiet first m.Run no longer reads survived. A TestMain that threw away a first
    m.Run's output left no fail event, and under -test.failfast the second m.Run ran nothing, so
    a mutant the suite catches read as a survivor. A green first run with a package that has tests and
    ran none now goes again without that flag, as one with a fail event already did. A target with no such
    package takes no extra run. New cost: any package under the target that passes and shows nothing of m.Run on a green run, including a TestMain that hides
    m.Run's output and calls it again only on failure and a TestMain gate that exits without m.Run, costs every
    survivor of the target one more suite run and one more first try for a flake, so a flake can still turn a true survivor there into a kill, exit 0.

  • The reruns without the mutant are no longer blind to a second m.Run. They kept
    -test.failfast, under which a TestMain that calls m.Run again exits 0 after a real failure,
    so a test that fails with no mutant at all read as a kill, witness kills included. They now run
    without that flag, and such a suite reads unconfirmed. No extra runs: a rerun that passes runs the
    same tests, and one that fails runs the rest of its package before the mutant ends.

  • A witness mark is good for one run of one mutant. The key lasted all of cross's run, so a
    signed line kept from one mutant's run and printed in another's read as a witness kill, under
    --require-witness, on a mutant no witness saw. Each go test command now signs with its own
    key, and such a line reads as self-report. So does one kept from a mutant's first run and printed
    on its repeat. No extra runs.

  • make test-db clears up after a killed run. A killed test binary left its --rm database
    containers running, and they held the host's AIO slots until MySQL could not start. Each
    container now names its owner in labels, and a run removes those whose owner is gone on this
    host, leaving any other run's alone.

  • subpoena-witness stops while waiting on a FIFO with no reader. The open came before the
    stall guard, so SIGTERM did nothing and only SIGKILL ended it. It now exits 1, with no start
    record and no ready file, as it did after SIGKILL.

  • A dangling link cannot put -out or -ready under a watched root. EvalSymlinks refuses a
    link to a file not made yet, and the check then passed. Links are now followed one at a time.

  • A TestMain that retries and shows its output no longer hides a kill. -test.failfast made a
    second m.Run a run of nothing that exits 0, so a suite that caught the mutant read as survived.
    A green run with a test's fail event in it now goes again without that flag. Red still has to
    hold as a kill, and a retry that truly passes is still survived.

  • A failed Example is a kill. It has a fail event and no file:line, so it read as a crash.
    Package testing's own got: report now counts as its failure message, as other evidence.

  • A printed log line after open text takes no witness kill. A line such as
    INFO server.go:42: handling request has the shape of a failure after text printed with
    no newline, and it read as the test's first failure, pushing a witness kill down to
    self-report. Such a line now counts only in a test with no other failure message, or when it
    carries the harness's mark.

  • A printed pass no longer turns a witness kill into unconfirmed. Code under test could print
    go test's framed --- PASS: line beside its real failure, and the test then read as passed. A
    witness failure with the run's mark now outlives a pass, unless another test started between
    the pass and the fail. A printed pass still costs a plain kill, and one in a test that then runs
    beside another costs a witness kill too: that reads unconfirmed, and the run stays red.

  • A fail printed after a real pass is no kill. Every test passed, TestMain exited 1 on the
    mutant, and text printed after the last test read as that test's failure. A pass ends a test,
    so only a witness failure's mark counts after one. The same goes for a fail printed for a test
    that never ran, in a package whose m.Run said PASS. A suite that hides that PASS by
    swapping os.Stdout around m.Run is not caught: such a fail still reads as a plain kill, as
    no output can show it, and only --require-witness stops it.

  • A second m.Run that runs nothing does not cost a real kill. Under -test.failfast it says
    PASS after a real failure, so the failure was dropped and the run read unconfirmed. Such a run
    now takes its failed tests from a run without that flag. Three shapes still lose the kill, loudly,
    reading unconfirmed: a second m.Run that truly passes; a failure only in the second m.Run,
    a signed witness failure included; and two m.Run calls with a later test that panics on the mutant.

  • Under --require-witness, a crash line is a crash in any test. A framed === NAME line could
    move a recovered panic's line onto a test that passed, and a witness failure then read as a
    witness kill. A passing test that prints such a line beside a witness failure now reads crashed.
    Without the flag the moved line is still a log line, and that plain kill stands: a plain kill
    assumes the suite is not forging test-runner output. See docs/architecture.md.

  • A printed pass does not hide a crash line. A test that recovers a panic stays a crash, as
    Ginkgo's does, and a printed pass switched that off for a test that really failed. A crash line
    now counts in any test with a fail event, whatever pass sits beside it.

  • A fail printed in a green package is no failure. A package counts only if its binary exited
    1, so such a line no longer carries a kill for a run that is red because another package crashed.

  • A package that hangs no longer costs another package's kill. Every package runs now, and a
    timed-out run went straight to enforcement-mutant-timeout, with a clean failure in the
    output. A run with a clean failure in a package that finished now goes to the repeat, and a
    hang with none stays a timeout. If the repeat hangs too, each package with a failed test goes
    alone, so a hang in another package c...

Read more

Subpoena v0.1.0-alpha.5

Pre-release

Choose a tag to compare

@nickhenggeler nickhenggeler released this 10 Nov 00:30
bd3ebca

Tagged v0.1.0-alpha.5 and v0.1.0-alpha.6, on the same commit: the fixes and the new witnesses
landed interleaved, so no earlier commit holds the fixes alone. No Python release goes with these
tags. PyPI stays at 0.1.0a4 until a python/ tag is cut.

What is new:

  • Three witnesses: file writes, database commits (Postgres and MySQL, experimental) and
    delivery on an inbound socket.
  • A kill has to hold: it reruns on the whole suite, with and without the mutant, and a Go kill
    needs a failed test that links the mutated package.
  • Sinkhole can answer 200, and Marked tells one arrival from another.
  • Cross output prints upstream line numbers and counts markers against mutants.
  • Python verdicts: --test-helper declares test code, Mock.assert_* under pytest-mock is a
    kill, and production code run by --doctest-modules no longer counts as test code.

Stated limits:

  • Forged runner output can fool a plain kill. Only a verified witness mark under
    --require-witness holds against it (docs/architecture.md, "What forged output can do").

  • A witness mark does not hold against code that reads the key and signs its own mark. Do not rely
    on one against a suite written to cheat.

  • A Pass from the database witness is experimental, and its known gaps are listed in its entry below.

  • A poll inside an exec waits for the arguments. ProcWitness could land after an exec set the
    new task name and before the arguments existed, and it kept that name as an unidentified event for
    good. Under an Allowed list that went red with nothing wrong, in about 2 of 100 batches of the
    gateway suite under load. Such a poll now re-reads the process for up to 5ms, and takes the line if
    the name, binary and cgroup still match. If they do not, or the wait runs out, the name stays and
    the check goes red as before. A read that pairs the old name with the new line can still go red.

  • A run that cost a kill says which test failed in it. An unconfirmed kill showed the last 20
    lines of the red run and the mutant run's failing tests. In a long suite those lines are the slowest
    tests' passes, so the test that failed without the mutant could not be seen. Where that run has a
    failure to name, the output now starts with its failed tests and first failure. What counts as unconfirmed is unchanged.

  • A panic testify's suite recovered is a crash. It reports one through t.Errorf, with no
    line at column 0, so it read as a plain kill. test panicked: after a file and line, in a test
    that did not pass, now counts as a crash, as Ginkgo's [PANICKED] does. New cost: a failed test
    that logs those words reads crashed.

  • A recovered panic in a parallel test is a crash. Package testing's own frames can put the
    panic: line of one parallel test on another that passed, where it read as a log line and the
    test's plain failure as a kill, with nothing forged. A crash line on a passing test now counts if
    that test ran beside one that did not pass, read from each test's events from run, or cont if
    it paused, to its end, whatever slashes the subtests' names hold. That covers subtests started from
    goroutines, with three exceptions that stay plain kills and are stated in docs/architecture.md: one whose name goes on from a
    sibling's, which passes and ends while that sibling runs, one whose parent carries on and recovers the panic in its own
    body, and one that works and then calls t.Parallel. Sequential suites
    keep the log-line rule, unless a TestMain that calls m.Run again repeats their names. New cost: a passing test that logs such a line beside a real
    failure reads crashed, and the run fails.

  • A quiet first m.Run no longer reads survived. A TestMain that threw away a first
    m.Run's output left no fail event, and under -test.failfast the second m.Run ran nothing, so
    a mutant the suite catches read as a survivor. A green first run with a package that has tests and
    ran none now goes again without that flag, as one with a fail event already did. A target with no such
    package takes no extra run. New cost: any package under the target that passes and shows nothing of m.Run on a green run, including a TestMain that hides
    m.Run's output and calls it again only on failure and a TestMain gate that exits without m.Run, costs every
    survivor of the target one more suite run and one more first try for a flake, so a flake can still turn a true survivor there into a kill, exit 0.

  • The reruns without the mutant are no longer blind to a second m.Run. They kept
    -test.failfast, under which a TestMain that calls m.Run again exits 0 after a real failure,
    so a test that fails with no mutant at all read as a kill, witness kills included. They now run
    without that flag, and such a suite reads unconfirmed. No extra runs: a rerun that passes runs the
    same tests, and one that fails runs the rest of its package before the mutant ends.

  • A witness mark is good for one run of one mutant. The key lasted all of cross's run, so a
    signed line kept from one mutant's run and printed in another's read as a witness kill, under
    --require-witness, on a mutant no witness saw. Each go test command now signs with its own
    key, and such a line reads as self-report. So does one kept from a mutant's first run and printed
    on its repeat. No extra runs.

  • make test-db clears up after a killed run. A killed test binary left its --rm database
    containers running, and they held the host's AIO slots until MySQL could not start. Each
    container now names its owner in labels, and a run removes those whose owner is gone on this
    host, leaving any other run's alone.

  • subpoena-witness stops while waiting on a FIFO with no reader. The open came before the
    stall guard, so SIGTERM did nothing and only SIGKILL ended it. It now exits 1, with no start
    record and no ready file, as it did after SIGKILL.

  • A dangling link cannot put -out or -ready under a watched root. EvalSymlinks refuses a
    link to a file not made yet, and the check then passed. Links are now followed one at a time.

  • A TestMain that retries and shows its output no longer hides a kill. -test.failfast made a
    second m.Run a run of nothing that exits 0, so a suite that caught the mutant read as survived.
    A green run with a test's fail event in it now goes again without that flag. Red still has to
    hold as a kill, and a retry that truly passes is still survived.

  • A failed Example is a kill. It has a fail event and no file:line, so it read as a crash.
    Package testing's own got: report now counts as its failure message, as other evidence.

  • A printed log line after open text takes no witness kill. A line such as
    INFO server.go:42: handling request has the shape of a failure after text printed with
    no newline, and it read as the test's first failure, pushing a witness kill down to
    self-report. Such a line now counts only in a test with no other failure message, or when it
    carries the harness's mark.

  • A printed pass no longer turns a witness kill into unconfirmed. Code under test could print
    go test's framed --- PASS: line beside its real failure, and the test then read as passed. A
    witness failure with the run's mark now outlives a pass, unless another test started between
    the pass and the fail. A printed pass still costs a plain kill, and one in a test that then runs
    beside another costs a witness kill too: that reads unconfirmed, and the run stays red.

  • A fail printed after a real pass is no kill. Every test passed, TestMain exited 1 on the
    mutant, and text printed after the last test read as that test's failure. A pass ends a test,
    so only a witness failure's mark counts after one. The same goes for a fail printed for a test
    that never ran, in a package whose m.Run said PASS. A suite that hides that PASS by
    swapping os.Stdout around m.Run is not caught: such a fail still reads as a plain kill, as
    no output can show it, and only --require-witness stops it.

  • A second m.Run that runs nothing does not cost a real kill. Under -test.failfast it says
    PASS after a real failure, so the failure was dropped and the run read unconfirmed. Such a run
    now takes its failed tests from a run without that flag. Three shapes still lose the kill, loudly,
    reading unconfirmed: a second m.Run that truly passes; a failure only in the second m.Run,
    a signed witness failure included; and two m.Run calls with a later test that panics on the mutant.

  • Under --require-witness, a crash line is a crash in any test. A framed === NAME line could
    move a recovered panic's line onto a test that passed, and a witness failure then read as a
    witness kill. A passing test that prints such a line beside a witness failure now reads crashed.
    Without the flag the moved line is still a log line, and that plain kill stands: a plain kill
    assumes the suite is not forging test-runner output. See docs/architecture.md.

  • A printed pass does not hide a crash line. A test that recovers a panic stays a crash, as
    Ginkgo's does, and a printed pass switched that off for a test that really failed. A crash line
    now counts in any test with a fail event, whatever pass sits beside it.

  • A fail printed in a green package is no failure. A package counts only if its binary exited
    1, so such a line no longer carries a kill for a run that is red because another package crashed.

  • A package that hangs no longer costs another package's kill. Every package runs now, and a
    timed-out run went straight to enforcement-mutant-timeout, with a clean failure in the
    output. A run with a clean failure in a package that finished now goes to the repeat, and a
    hang with none stays a timeout. If the repeat hangs too, each package with a failed test goes
    alone, so a hang in another package c...

Read more

Subpoena v0.1.0-alpha.4

Pre-release

Choose a tag to compare

@nickhenggeler nickhenggeler released this 02 Nov 18:40
0594ede

Tagged v0.1.0-alpha.4 and python/v0.1.0a4. Hardens Subpoena's own checks;
no new language surface.

  • Cross --require-witness (Go and Python). Opt-in strict mode: a clean
    kill whose first evidence is not a witness becomes
    enforcement-mutant-killed-by-self-report and fails the run. Selection
    and group masking honor that outcome. The gateway and adopter examples run
    strict, with their witness assertions first; the Level 0 Python quickstart
    and cross-self stay in ordinary mode. The default is unchanged. (#19)
  • Lint resolves telemetry constants. Go unknown-telemetry now resolves
    constant values with go/types (lexical fallback for unchecked files), so
    a constant can no longer hide an unknown key. Both Level 0 quickstarts are
    linted in CI. Vocabulary exemptions need a printed reason and must
    suppress a real finding; unwitnessed enforcement can be exempted only by a
    reasoned marker on one function. Missing, misplaced, duplicate or stale
    markers, and unused invocation exemptions, exit 2 in both linters. (#18)
  • Tier 2 finalizes kernel capture before OATS judges absence. A
    fixture-only controller (conformance/oatscapture) closes and drains the
    witness, waits for it to exit and validates the complete log before the
    absence assertions run. Both cases also require one start, one stop
    and a healthy, error-free stop record in Loki. Healthy kernel capture does
    not imply lossless collector or Loki delivery. (#17)

Subpoena v0.1.0-alpha.3

Pre-release

Choose a tag to compare

@nickhenggeler nickhenggeler released this 01 Nov 12:47
0207ff0

Tagged v0.1.0-alpha.3 and python/v0.1.0a3. Python gains network
absence, lint, cross-examination with layer groups, session cgroups, the
production exporter, invariants by name and a pytest plugin. Running a
witness and registry-drift stay Go only.

  • Python pytest plugin. Installing subpoena registers a pytest11
    plugin with subpoena_harness, subpoena_watch_spawns and
    subpoena_sinkhole fixtures, so a test no longer wires Harness,
    addfinalizer or the sinkhole's teardown by hand.
  • Python network absence (W6). subpoena.witness.Sinkhole is the
    loopback network witness: it records every connection that reaches it,
    drains its accept queue on stop(), and reports health only once stopped.
    subpoena.harness gains no_dial, require_no_dial, check_no_dial and
    dial_to, under the same verdict rules as Go's NoDial, plus
    check_no_spawn. docs/ports.md makes network absence part of the port
    contract.
  • Python invariants by name. subpoena.invariants.load(name) reads
    invariants/<name>.traceql from the nearest invariants/ above the
    caller (or a given directory), as Go's invariants.Load does; an unknown
    or path-shaped name raises InvariantError. names() lists them.
  • Python Trace.query returns a Result with matches, complete and
    count(), as in Go; iterating or len() still goes over the matches.
    Trace.spans(trace_id) returns one trace's spans.
  • Python session cgroups (W2). subpoena.cgroup ports pkg/cgroup, so
    a Python service can run a session's processes in
    <root>/<trace id>, the cgroup the witness attributes by. On a real cgroup
    the child places itself between fork and exec (Python has no
    CLONE_INTO_CGROUP). Otherwise the PID is written after start, and a
    failed placement comes back as place_err so the caller can fall back to
    lineage.
  • Python production tiers. subpoena.production ports pkg/production: Exporter wraps a span exporter and removes test-tier attributes from spans, events and links without mutating the originals; collector_config emits the same collector transform processor as Go. test_tier_keys lists the test-tier keys.
  • Python lint (L1). subpoena-lint <dir> (console script,
    subpoena.lint) runs the eight internal/lint rules over Python source
    with ast, resolving imports and aliases, so a renamed require_absent
    is still the harness call. Unreadable files, invalid Python, missing
    explicit configuration and star imports fail the run with exit 2.
    Registry-drift stays a Go-only subcommand.
  • Python cross-examination catches up with Go. Layer groups with a
    reviewed cross-masked.txt, parallel mutant runs (-j), -select with a
    clean-baseline fallback to the full suite, and a per-result evidence file
    under -evidence-dir. A grouped survivor never counts as killed, and a
    masked layer stays listed.

Subpoena v0.1.0-alpha.2

Pre-release

Choose a tag to compare

@nickhenggeler nickhenggeler released this 29 Oct 02:38

Tagged v0.1.0-alpha.2 and python/v0.1.0a2. Python gains the witnessed
absence checks and cross-examination; lint is still Go only.

  • Python cross-examination (G1). subpoena-cross <dir> (console script,
    subpoena.cross) forces each # subpoena:enforcement_guard condition and
    replaces each # subpoena:enforcement statement with pass, in a copy, and
    requires pytest to fail on an assertion raised from test code (a test
    module, a conftest, or a subpoena check such as watch_spawns). An assert in
    the code under test, other exceptions and dead test processes are crashed;
    a mutant that does not import, a collection error, or a module shadowed by an
    installed package is invalid; a hang or a pytest-timeout failure is
    timeout. All fail the run, as does any marker that takes no effect. The
    point marker is a full-line comment directly above the def, as Go reads a
    doc comment. Setup faults (a tree that cannot be copied, non-UTF-8 marked
    source, a copy that drifted from the discovery snapshot) exit 2, and Ctrl-C
    kills the running suite. examples/python-quickstart is now
    cross-examined in CI. Groups, cross-masked.txt, -j and -select are Go
    only for now.

  • Witness file health (W4), and absence checks in Python (W5).
    subpoena-witness writes a start record once attached and a stop
    record with the capture's health (lost kernel events, read or write
    failures), which it used to drop. witness.Log now reports Health, so
    NoSpawn over a witness file can Pass; before, it was always Inconclusive.
    A file with no records, more than one capture window (nothing watched
    between them), an exec outside the window, or an unhealthy or missing stop
    is a capture fault. A malformed line, a time without a zone, NaN, or a key
    that matches a field only by case is a read error, in Go and Python alike. Python gains the same
    Log.health(), no_spawn, require_no_spawn, watch_spawns and
    SpawnRule, and ProcessEvent.name_only. conformance/witness.yaml holds
    the cases both readers pass.

  • Cross: evidence per kill, and crashes. Each killed mutant reports its
    first evidence, from the first failure message of a failed test: witness,
    testimony (the harness tags) or other. A suite that goes red only by a
    crash (a panic, a runtime fatal error, os.Exit, or a failure with no
    message) is enforcement-mutant-crashed and fails the run. -evidence-dir keeps
    every mutant run's full go test -json output.

  • Cross: -j N parallel mutants. Single mutants run N at a time, each in
    its own copy of the module. cross-self: 490 s at -j 1, 135 s at -j 8, same
    verdicts. Load can flip a verdict either way, so gate on -j 1.

  • Cross: -select test selection. Each mutant runs only the tests that
    executed its line in a per-test coverage pass after the baseline. Opt-in: it
    pays off on long suites, not on short ones. A mutant its selected tests do
    not kill runs again on the whole suite, and group runs use the whole suite. A
    test that fails alone, or a selected subset that fails together without any
    mutant, is an error.

  • Guard and layered mutants in subpoena-cross.
    //subpoena:enforcement_guard true|false on an if forces its condition to
    the bypass value. group=<name> on enforcement markers declares layers of one
    guarantee: cross also mutates the whole group, and a layer that survives alone
    but is caught with its group is enforcement-layer-masked, naming the masking
    layer. A masked layer fails the run unless cross-masked.txt lists it with a
    reason; a stale entry fails too.

  • Breaking: enforcement markers no longer accept trailing text.
    //subpoena:enforcement because ... is now a discovery error; only
    group=<name> may follow the marker. A marker that would take no effect (a
    statement or guard marker outside an enforcement point, or a misspelt one) is
    an error too; before, it was ignored and its statement went unexamined.

  • Cross run hygiene. A timed-out mutant names the packages and tests still
    running at -timeout. Ctrl-C (or SIGTERM) cancels the run, killing the
    mutant's test processes and removing the copies. Where a timeout cannot kill
    the test binary (not unix), the run stops at the first timeout. A guard
    marker on a line whose outermost statement is not the if (an if inside
    an inline closure) is an error.

  • Cross refuses ambiguous cross-masked.txt entries. One entry accepts one
    masked layer; an entry matching two, or listed twice, is an error. An entry
    names the layer's group (<file> | <func> | group=<group> | <removed text> | <reason>), so a layer moved to another group needs a fresh review. A canceled
    run is an error, not a kill.

  • harness.WatchSpawns (Go). A witness check registered when the witness
    starts that runs at test cleanup, even after an earlier Fatal: it stops the
    witness, fails on commands the SpawnRule refuses (Allowed, Never) and on
    missing or bad capture health, and logs every process seen when the test had
    already failed. A Never entry's program
    matches with or without a directory (sleep 1 refuses /usr/bin/sleep 1).

  • harness.CheckNoSpawn, CheckNoDial (Go). Non-fatal forms of the witness
    absence checks. subpoena-lint counts them, and WatchSpawns, as asking a
    witness.

  • Name-only sightings are marked. When ProcWitness sees a process only
    by its kernel task name (it exited before its arguments were read),
    ProcessEvent.CommandSource is comm and NameOnly() is true. WatchSpawns
    fails on a name-only sighting whose name is a Never program, and, with an
    Allowed list, on any name-only sighting unless SpawnRule.ExcuseNameOnly
    is set and the name is an Allowed program. Before, a bare name could slip
    past Never.

  • Tagged failure messages. Harness witness failures start with witness:,
    telemetry assertion failures with testimony: (Go and Python).

  • Lint: verified Collect wrappers. A test helper that stops the system
    and witness, checks health and calls the harness Collect counts as
    Collect for no-trace-assertion and unguarded-absence, but only when
    every return hands back the trace from that call. A same-named method that
    doesn't, such as a metrics Collect, still doesn't count (lint/README.md).

  • ProcWitness.Coverage (Go). The number of full scans and children walks, and
    the longest window in which a process outside the root's children tree could
    have started and exited unseen (from one scan's start to the next one's end,
    or to Stop). A process shorter than that window may have been missed. CI now runs the BenchmarkProc* benchmarks (report only).

  • Python harness: span limits (fix). The Python harness now sets its own
    span limits, all unlimited, so OTEL_*_LIMIT variables no longer drop or
    truncate attributes (H1). collect() raises IncompleteError (.truncated)
    when a span lost attributes, events or links (H3). In 0.1.0a1, a span with
    more than 128 attributes silently lost its oldest ones, and
    OTEL_ATTRIBUTE_VALUE_LENGTH_LIMIT truncated values. Both still reported a
    complete capture. conformance/harness.yaml gains an env field and an
    env_limits_ignored case, and the Python runner now applies filler.