25 KiB
Framing — group cleanup fails silently at four sites
Revision 4. Status: APPROVED at revision 3; implemented. Lane
reap-ledger-silent-failures, worktree ../pmacs-reap-ledger, based on
githubsucks/main @ 22df6ab. Revision 4 records what implementation
found; it is not a new design round.
Parked by PR #200's framing §5 and unparked by its evidence. #200 retired the premise that justified the ledger's leniency; it deliberately changed no disposition, and said so. This lane owns what it refused.
Revision history
Revision 3 → 4, found while implementing, not a new design round. Every bet resolved; one acceptance turned out to be satisfiable vacuously.
- Acceptance 2's in-drain clause could be met by a vacuous fixture,
and was. "The live descendant's named late output absent" says
nothing about how the descendant stays live — and
poll_onesendsSIGTERMto the whole group on leader exit, so an untrapped descendant dies before it can write. The marker was then absent on both paths and the pin would have stayed green with the collapse fixed. The bite is what caught it: with the seam reverted, the pin failed only the consumed-plan check, never the content assertion. The fixture now usestrap '' TERMbehindsurvivor_script's readiness gate. The lesson generalises past this pin: an absence assertion is only as good as the fixture's ability to produce the thing. - Bet 1 holds. All four sites took a directed outcome with no
restructuring.
final_drain_runtime— the one §3 named as at risk, being a free function — needed only a shared handle on the context it already receives. - Bet 2 holds, in the direction that keeps the lane. All four consequences are reachable; none was already foreclosed by an earlier guard. The lane does not shrink.
- Bet 3 resolves: the coupling is real and measured. With a failed
force-kill and an errored probe,
shutdown()returns in under 500ms instead of holding its 2s bound, with the survivor alive. With only the failed force-kill it burns the full bound. §1.3's warning stands: making the probe strict without touching the loop converts a silent early exit into a guaranteed 2s stall. - Bet 4 is falsified, exactly as its own clause anticipated: no
channel exists.
ProcessEventis keyed byProcessIdwhile the ledger is keyed by pgid and is deliberately independent of managed records, so in the leader-exited-survivor case there is no id to attribute to. Every production consumer pollstake_events(id)per known id (lua_bindings/mod.rs:8933,:10731,mcp.rs:361).take_all_eventswould sidestep the keying but has no production consumer at all — its only two call sites are tests, despite a doc comment naming a*processes*buffer.pmacs.errorwas already known dead. Reporting becomes its own lane (§5), and this PR ships instrumentation plus tests without it, which is what §7 said it would do in this case. - The in-drain
SIGKILLis still unpinned, as §1.2a promised. Itsgroup_killedflag stays local and the persistent ledger retries in the same outer tick, so pinning the local non-retry still needs a call-count assertion or a direct free-function test. Unchanged, and restated here so a later reader does not mistake the shipped seam for covering it.
Revision 2 → 3, after review round 2 (two blocking, two major). All four accepted; all four verified in the code first.
- The new in-drain test could not fire under the specified seam. The
drain initializes
last_dataon entry and probes again after each 1 ms sleep;quiescedneeds a false result for the full 50 msREADER_SEND_POLL_INTERVAL. A one-shot error is therefore gone before reader cancellation. The in-drain override now returns its selected result for that oneGroupDrainCtx, and the acceptance requires named late output to prove the cancellation throughpoll_one. - The in-drain SIGKILL had no outer-path discriminator. Its local
group_killedflag is followed by the persistent ledger's retry in the same outer tick; testing it would require a call-count assertion or a direct free-function test. It remains an explicit code fact but is outside this diagnostic PR's injection seam. - The seam needed a lifetime owner. A context-local queue cannot be
asserted at fixture teardown and a global one crosses test fixtures.
The test state is now specified as per-supervisor and shared into the
production
GroupDrainCtx; an unconsumed outcome proves the intended site was never reached. - The four-site scope had stale "three" language. Acceptance and handoff obligations now distinguish the three persistent-ledger paths from the in-drain probe twin.
Revision 1 → 2, after review round 1 (three blocking, two major). All five accepted; all five verified in the code first.
- §0 named the wrong journey step. It claimed step 8, "Open a
terminal", while Q#RL5 in the same document says the ledger is
unreachable for PTY children. Both cannot be true. The only production
group = truecaller is compile mode (builtin/runtime/compile.lua:820), so this is step 9, build/test, plus general pipe-process cleanup. - A fourth site was missed.
final_drain_runtime(:2331) probeskill(-pgid, None), treats every error as dead, discards itsSIGKILLresult and sets its owngroup_killedflag — and it enforces the ledger deadline while no tick runs. It is now in scope (§1.2a). - The staging contradicted itself. "Bet 1 ships alone" against an acceptance requiring three failure-path tests, under one-branch / one-PR. A seam with no tests does not even prove it reaches the production calls. §7 now scopes the first PR as seam plus behaviour-preserving tests, which is still diagnosis-only.
- A generic seam is the wrong shape.
shutdown()callsself.signal(*id, SIGKILL)before its ledger force-kill, so a single "next kill errno" would be eaten by the wrong call. The seam is now site-directed (§3 Bet 1). - §1.3 overstated the loop coupling. Early exit needs the ledger
empty and
any_running()already false; a live managed record keeps the loop going regardless. The precondition is now stated.
Diagnosis first. No disposition change is proposed in this revision. §2 asks whether one is warranted; §5 parks every candidate until the lane's own evidence exists. That ordering is not caution for its own sake — Stage A of the signal lane had three tolerance rules rejected across three revisions, each because it concluded something about one entity from something that was not about that entity, and this is the same shape of problem on the same data structure.
0. Coherence impact (COHERENCE §20)
- Journey step 9, "Build and test" — every compile and grep run,
plus general pipe-process cleanup. Not step 8: the ledger arms
only for
proc.spec.group, which spawn rejects for PTY mode, so no terminal ever reaches it (Q#RL5). The only productiongroup = truecaller is compile mode (builtin/runtime/compile.lua:820). No grade change proposed. - Serves §9 (worker model), failure attribution. The ledger is the one mechanism that can see a survivor nothing else can, and today it cannot report that it failed to.
- Interaction islands: none. Config registry: not adopted. Background-work attribution: unchanged.
- No audited claim in COHERENCE.md changes; under §25 no COHERENCE edit rides this PR.
1. Ground truth (verified at 22df6ab)
1.1 What the ledger is for
tick_reap_ledger (src/process.rs:1472) exists for one case its own
doc names: a TERM-ignoring descendant that survived its leader's clean
exit with its output redirected. Neither leader state nor reader state
can see that survivor — try_wait reports the leader, and the readers
see a closed pipe. Only group liveness can (Q#CM3, round-3 finding 1).
That is the blast radius. A silent failure here leaks precisely the process the mechanism exists to catch, and nothing else in the supervisor is looking.
1.2 Three silent failures in the persistent ledger
// (a) any probe error drops the entry
if nix::sys::signal::kill(Pid::from_raw(-*pgid), None).is_err() {
return false;
}
// (b) the escalating SIGKILL's result is discarded, and the entry is
// marked killed regardless
if now >= entry.deadline && !entry.killed {
let _ = nix::sys::signal::kill(Pid::from_raw(-*pgid), Some(Signal::SIGKILL));
entry.killed = true;
}
(a) retain returning false deletes the entry, so escalation is
cancelled. The comment is honest that this is a bounded-growth policy
rather than a claim the group is gone — #200 corrected it — but the
behaviour is unchanged: an EPERM probe is indistinguishable from
ESRCH.
(b) A failed SIGKILL is recorded as a successful one. The entry
then satisfies !entry.killed == false forever, so no later tick
retries it — the escalation arm is guarded by !entry.killed and
never fires again for that group.
Scoped to ticks, and the scope matters. shutdown()'s force-kill
loop (c) iterates the ledger with no !entry.killed guard, so it
does re-kill an entry this arm marked. The two failure modes are
therefore distinct rather than cumulative:
| Failure | Survivor lives until |
|---|---|
escalation SIGKILL fails |
editor exit, where shutdown() gets one more attempt |
shutdown() force-kill fails |
past editor exit — nothing else tries |
Saying (b) is "never retried by anything" would collapse that distinction and overstate it: the one remaining attempt is exactly what (c) is, and (c)'s own failure is a different and worse outcome.
(c) shutdown() has the same discard (:1763-1766), on the path
that exists specifically to stop a leak at editor exit:
for (pgid, entry) in &mut self.reap_ledger {
let _ = nix::sys::signal::kill(Pid::from_raw(-*pgid), Some(Signal::SIGKILL));
entry.killed = true;
}
Its own comment says it is there because "a pre-deadline ledger ... would be silently discarded at Drop and leak the member". The fix for one silent leak was written with a discarded result of its own.
1.2a A fourth site: the in-drain twin
final_drain_runtime (:2331, called once from :1573) runs the same
pattern on the same pgid, while no tick is running:
let group_alive = nix::sys::signal::kill(Pid::from_raw(-ctx.pgid), None).is_ok();
if group_alive && now >= ctx.deadline && !group_killed {
let _ = nix::sys::signal::kill(Pid::from_raw(-ctx.pgid), Some(Signal::SIGKILL));
group_killed = true;
}
is_ok() collapses every errno into "dead", exactly as the persistent
ledger's is_err() does — and the consequence differs. Once that false
"dead" persists for one quiescent READER_SEND_POLL_INTERVAL, it makes
quiesced true, which sets rt.cancel and cancels the readers, so
the failure mode is truncated output rather than a leaked process. One
false probe is not enough: the loop probes again every millisecond, and
last_data starts at the drain's entry.
It is not identical to the persistent ledger and the framing does not
claim it is. A later tick can retry the ledger entry; this decision
is terminal for that drain. It is in scope because it is the same
collapse on the same data with its own consequence, not because it is
the same bug.
The ignored in-drain SIGKILL is a real code fact, but not an
independently observable acceptance in this diagnostic PR. Its
group_killed flag is local; when the drain returns, the same outer
tick reaches the persistent ledger and can retry the group. Proving
the local non-retry without a call-count assertion would mean testing
the free function directly rather than its production caller. This lane
therefore tests the probe-collapse consequence above and does not
promise an injection seam for the in-drain SIGKILL. Any later retry
policy must re-scout that local flag with the persistent ledger.
It constrains the seam. final_drain_runtime is a free function
taking &RuntimeHandles and Option<GroupDrainCtx> — there is no
&mut self to hang a supervisor field on. Test-only injection state
therefore remains owned by the ProcessSupervisor and is shared into
GroupDrainCtx at the :1573 production call site. It is never global:
fixture teardown can then assert the planned outcome was consumed, even
when unit tests run in parallel.
1.3 The shutdown loop's exit condition depends on the silent drop
shutdown()'s final loop (:1773-1775) runs while
self.any_running() || !self.reap_ledger.is_empty(), bounded at 2s, and
calls tick() — which calls tick_reap_ledger.
So the loop terminates when the ledger empties, and the ledger empties
via (a). On ESRCH that is correct: the group is gone. On any other
errno the loop exits early, having concluded cleanup finished because
the probe failed.
The precondition matters and revision 1 omitted it. The condition is
a disjunction: any_running() || !reap_ledger.is_empty(). An early exit
therefore needs the ledger empty and any_running() already false —
a live managed record keeps the loop running whatever the ledger does.
The coupling is real for the case the ledger exists to serve, the
leader-exited survivor, and Bet 3 must build exactly that fixture rather
than any group.
This coupling is the reason (a) cannot be changed casually: making the probe strict without touching the loop converts a silent early exit into a guaranteed 2-second stall at every editor exit that hits it.
1.4 None of the three has been observed
#200 observed an explicit SIGTERM failing in signal(), not any
ledger call. What it established is narrower and still sufficient to
open this lane: a group-directed kill computed from the spawn-time
pgid == pid assumption returned EPERM while the leader was alive, on
macOS, intermittently
(run 30553376486).
That retires "EPERM cannot happen for our own children" as a reason to discard an arbitrary group error. It does not show that any of (a), (b) or (c) has fired. This lane must not claim otherwise, and §3's bets are built to find out rather than to assume.
1.5 The ledger's kills have no injection seam
signal() consults forced_kill_errno (Q#PD4), which is how #176 and
#200 tested failure paths without provoking real errnos.
tick_reap_ledger calls nix directly and consults nothing. Same
for shutdown()'s loop.
So all three failure paths are, today, untestable. That is the first thing this lane has to fix, and it is a prerequisite for any disposition change rather than a nicety: a disposition change whose failure path cannot be exercised is a rule nobody can falsify.
1.6 What is tested today
liveness_probe_reaps_term_ignoring_survivor_after_leader_exit
(:4274), repeated_terminate_does_not_extend_ledger_deadline
(:4314), shutdown_force_kills_outstanding_ledger_groups (:4350),
leader_exit_reap_bounds_drain_with_pipe_holding_descendant (:4415),
and setsid_escapee_is_not_reaped_and_teardown_reclaims_readers
(:4450).
Every one exercises the success path. None injects a probe or
SIGKILL failure, because §1.5 makes that impossible.
repeated_terminate_does_not_extend_ledger_deadline is also the test
that flaked on macOS in #191 — the occurrence that produced §1.4's
evidence. It is load-sensitive and unrelated to what this lane changes;
noted so a red run on it is not mistaken for this lane's doing.
1.7 Limits of the evidence
- Not reproduced. No ledger probe or
SIGKILLhas been seen to fail, on any platform. - The mechanism is not established. Why a group-directed
killcan returnEPERMagainst a live owned child is still unknown — #200 narrowed it and explicitly did not solve it. - Group identity remains unprovable (#200 §1.5). Nothing this lane
measures can distinguish the original group from a recycled one; a
pidfdcovers a process, not a group, and macOS has neither.
2. Questions
- Q#RL1 — Should the probe distinguish
ESRCHfrom other errnos? Unknown, and deliberately not proposed yet. It is the obvious change and it directly trades bounded growth (the stated original reason) for correctness, while §1.3 shows it also changes editor-exit timing. It needs evidence and its own review. - Q#RL2 — Should a failed escalation be retried, and how many times?
Unknown.
killed = trueon a failedSIGKILLis clearly wrong as bookkeeping; what should replace it is a policy question, not an obvious fix. - Q#RL3 — How does a background tick report anything? ANSWERED at
revision 4: it cannot, today.
ProcessEventis keyed byProcessId; the ledger is keyed by pgid and is deliberately independent of managed records, so in the leader-exited-survivor case — the one the mechanism exists for — there is no id to attribute to. All three production consumers polltake_events(id)per known id (lua_bindings/mod.rs:8933,:10731,mcp.rs:361).take_all_eventswould sidestep the keying but has no production consumer; its only call sites are two tests, notwithstanding a doc comment naming a*processes*buffer.pmacs.errorwas already dead at 15 sites. Reporting is therefore parked as its own lane (§5). - Q#RL4 — Does the
shutdown()loop need its own termination condition if (a) becomes strict? Almost certainly (§1.3), and that coupling is why the two cannot be changed independently. - Q#RL5 — Is the ledger reachable for PTY children? No —
spec.groupis rejected at spawn for PTY mode, so the ledger is a pipe-path mechanism only. Stated because #200 revision 2 got this exact relationship wrong in the opposite direction.
3. Bets
-
Bet 1 — the four sites can be made injectable, site by site, without changing behaviour. Not a generic "next kill errno":
shutdown()callsself.signal(*id, SIGKILL)before its ledger force-kill, so a single shared one-shot would be consumed by the wrong call and the test would pass while proving nothing.The persistent-ledger and shutdown outcomes are directed and one-shot. The in-drain probe instead needs a directed result for one complete drain: a one-shot error is consumed by its next 1 ms probe and cannot reach
quiesced's 50 ms interval. The test mode must therefore return the selected error for every in-drain probe until thatGroupDrainCtxends, while leaving the later persistent-ledger probe real.The independently addressable outcomes are:
Site What is injected tick_reap_ledgerprobethe kill(-pgid, None)resulttick_reap_ledgerescalationthe SIGKILLresultshutdown()force-killthe SIGKILLresultfinal_drain_runtimeprobeone selected result for the complete drain, via GroupDrainCtx(§1.2a)Bet 3's coupling test needs two at once — a failed shutdown force-kill and a failed subsequent probe — so the seam must express more than one pending outcome. A typed queue per site, or a per-site slot, either is acceptable; a single global slot is not.
Fixture cleanup is part of the seam, not an afterthought. Test state is per-supervisor, not global. An unconsumed planned outcome proves the fixture missed its intended production site; each finite outcome is consumed on use and asserted empty at fixture teardown, and the in-drain override records that it was used before its context ends.
- Falsified if a site cannot take a directed outcome without
restructuring the code under test — which would make the test a test
of the restructuring.
final_drain_runtimeis the one at risk, being a free function.
- Falsified if a site cannot take a directed outcome without
restructuring the code under test — which would make the test a test
of the restructuring.
-
Bet 2 — each silent consequence is demonstrable once injectable. With the seam: an
EPERMprobe drops an entry whose group is still alive; a failedSIGKILLleaveskilled = trueso no later tick retries it;shutdown()'s discard leaks the group past editor exit, which is a different and worse outcome; and a continuously false in-drain probe cancels readers before a live descendant's deliberately late output can arrive.- Falsified if any of the three turns out to be unreachable in practice — for instance if some earlier guard makes the entry already absent. That would be a genuinely good outcome and would shrink this lane rather than embarrass it.
-
Bet 3 — the shutdown coupling is real and measurable. A test shows the final loop exiting early when the probe errors, rather than running to its 2s bound.
- The fixture must have
any_running()already false (§1.3): a leader that has exited leaving a group survivor. Any other shape tests the disjunction's other arm and proves nothing about the coupling. - Falsified if the loop still runs to its bound with the ledger emptied and no managed record live — in which case §1.3 overstates the coupling and the two changes can be separated after all.
- The fixture must have
-
Bet 4 — a failure here is reportable at all. Q#RL3 has no answer yet. This bet is a scouting obligation, not a design: find every channel a background tick could use, and say plainly if none exists.
- Falsified if the only available channel is one already known dead
(
pmacs.error) — in which case reporting becomes its own lane and this one ships instrumentation plus tests without it.
- Falsified if the only available channel is one already known dead
(
4. Acceptance
- A directed, multi-outcome injection seam covering the three
persistent-ledger paths plus the in-drain probe of §3 Bet 1, on
Q#PD4's terms: result only, everything else production code. The
in-drain result lasts only for its one production
GroupDrainCtx. Finite outcomes are consumed on use and asserted empty at teardown. - A test per silent consequence, each asserting the observable consequence
— entry dropped while the group lives;
killedset after a failed kill; the same at shutdown; readers cancelled after a false "dead" in the drain and the live descendant's named late output absent — rather than that a function was called. - Each new test falsified by an actual revert, both directions recorded in the PR body.
- The
shutdown()coupling of §1.3 pinned by a test, whichever way Bet 3 resolves. - No disposition change.
retainstill drops on any error;killedis still set unconditionally. This lane makes the failures visible and testable; changing them is §5's. - Q#RL3 answered in the PR body with the channels actually found, or an explicit statement that none exists.
docs/agent-handoff.mdrecords the three persistent-ledger paths and the in-drain probe twin, and that none has been observed — the distinction #200 had to make twice.
5. Parked
- Q#RL1's strict-
ESRCHprobe, Q#RL2's retry policy, and any other disposition change. Each needs this lane's evidence. - Reporting, if Bet 4 finds no channel.
- Retargeting to the measured pgid, and any EPERM/ESRCH tolerance
rule in
signal()— still parked from #200, unchanged. signal_target's read-then-kill oftcgetpgrpon the PTY path — Stage A's "most likely real fix site", still unframed, still not this.- Why a group-directed
killcan returnEPERMagainst a live owned child. The mechanism is unknown; this lane instruments the consequences, not the cause.
6. Gates
Standard suite, each its own step with a real exit status and nothing
after the command that could mask it: cargo fmt --check; cargo clippy --workspace --all-targets -- -D warnings; cargo test --lib; cargo test --lib --features crdt; compile_mode_acceptance;
terminal_copy_mode_acceptance (both feature configurations); cargo test --test m4_acceptance -- --skip basedpyright;
PMACS_REQUIRE_GPU=1 cargo test -p pmacs-gpu; git diff --check.
All five bootstrap-storage variables controlled locally —
XDG_CONFIG_HOME, XDG_DATA_HOME, XDG_STATE_HOME, XDG_CACHE_HOME,
PMACS_STATE_HOME — per the ambient-root framing merged as #201.
Isolating only the config root leaves the data-root write path open.
Process-supervisor tests are load-sensitive; the PR body records repetition counts rather than a single green.
7. Branch plan
One branch, one PR: the seam together with the tests that exercise it. Revision 1 said "Bet 1 ships alone", which contradicted an acceptance requiring failure-path tests under one-branch/one-PR — and was wrong on its own terms, because a seam with no tests does not even demonstrate that it reaches the intended production calls.
This is still diagnosis-only: every test pins current behaviour, including the behaviour that is wrong. Nothing in this PR changes what the supervisor does.
If Bet 2 finds a path unreachable, the lane shrinks and says so rather than manufacturing a failure to justify itself. If Q#RL3 finds no reporting channel, reporting becomes its own lane and this PR ships instrumentation plus tests without it.