MobileGL/MobileGL/MG_Remote/CONTRACT-P6.md

CONTRACT-P6 — the backend runs in a second process

Authority: this file, beside CONTRACT-P5.md (table 0, byte carriers, field ownership, R-1…R-17), CONTRACT-P5B.md (the class-C slots), CONTRACT-P5C.md (rule E, the two named exemptions, SEG_EVENT, the guards), CONTRACT-P5E.md (rule F, the barriered predicate, the wait rule) and CONTRACT-MAGMA-RUNAHEAD.md. Where it disagrees with any of them this file is newer and wins; §10 lists every such place. Base: feat/disaggregated @ c25a7760. Every file:line was read at that commit; paths are under MobileGL/ unless they start with docs/.

How to change it. Package c6's file, edited by the integrator first. A P6 package (lk / hs / so / sm / cp / dl / st / t6) that needs a row changed goes through the integrator; the packages compile against the rows below from day one.

What this file is written from. a6-audit-v1.md (198 rows, seven items, each adversarially re-resolved) and a6-link-experiment.md (the 184 symbols). It supersedes P6-CONTRACT-DRAFT.md, which was written before its own audit and is wrong in four places (§10.1). The re-scoping against the eventual client-in-a-VM end state is P6-ENDSTATE-REVIEW.md; this file takes from it only what P6 must not preclude, and §11 says what P6 explicitly does not take on.

---

§0 What c6 is, and the rule above every row

c6 lands this file and two headers. It changes no behaviour, adds no caller, and moves no byte on any wire. Every rule below that describes an action names the package that implements it. A rule with no package named is a statement about the tree as it already is.

Rule G — a role may not name a process-local handle. Rule E forbade naming the other role's memory; rule G extends it to identity. A value whose meaning comes from the owning process's kernel or loader state — EGLDisplay, EGLSurface, EGLContext, ANativeWindow*, any void* window handle, an fd number, a pid, a mapped address — either crosses as a token defined in §2, or does not cross. Descriptors cross only by a mechanism the transport declares it supports, and where the transport declares none they do not cross at all and the bytes are transferred instead; a transport may not invent a third answer.

The draft hard-coded SCM_RIGHTS inside rule G. That is one transport's implementation of one clause, not the rule, and writing it into the rule is what would make the rule wrong the first time a transport has no descriptor passing at all.

Rule H — a value whose meaning comes from the peer's window system never crosses. Permanently. Not "until P12". P12 gives the server its own display; it does not give the client's window a way across. §6 is rule H's whole enforcement.

Rule G's teeth are §9's arm-proof gate: a spawn lane that silently ran monolith satisfies every other gate in this file.

---

§1 The one fact that makes c6 possible: Spawn is unreachable today

ConfigLoader.cpp:339-346 is the only writer of MG_Config::Transport for the string spawn, and it writes TransportMode::Monolith at :344, with a log line saying P5 does not implement it.

Two consequences, and the second is why this file can exist without changing behaviour:

1. ConfigLoader's named refusal cannot falsify itself. It lands on the same value a server's resting state would have, so "the child was forced to monolith" and "the parent fell back" are one state. sm fixes this; until then, no lane can prove it ran spawn (§9.5).
2. Every rule in this file conditioned on Transport == Spawn is vacuous in c6's own tree. It constrains a state no process can enter. That is how D1 (§3.1) is stated normatively while changing nothing, and c6's gate pins the vacuity so the ConfigLoader change cannot land early.

---

§2 Table 0 additions

row	encoding	invalid	consumer	package
DisplayToken / SurfaceToken / ContextToken	Uint64, minted by the client, dense from 1, never reused in a session. Server keeps token → native handle; it never sees the client's EGL values. protocol.fbs's display / surface / context are already ulong and are these. The client already mints them (MG_State/EGLState/Core.h's EncodeHandle), contrary to Server/SurfaceControlFrame.h:26-29, which says inproc bit-casts the handle — that comment is corrected by cp.	0 on any op but ReleaseCurrent → Fatal{ProtocolCorruption}	the server's surface map	cp
SurfaceOpKind::InitCapabilities = 11	Appended. The draft claimed ServerInitCapabilities needs no wire form because "CapsSnapshot answers it". The answer is a CapsSnapshot; the request has a live client caller (Client/BackendObject_Remote.cpp:126) and the codec refuses to encode it (Protocol/SurfaceOpCodec.cpp:119-122, InprocOnlyOpOnTheWire). A spawn client cannot ask.	unknown tag → Fatal{ProtocolCorruption}, never ignored	the server's control pump	cp
SurfaceReply.result: int	Appended, id 5. Carries the MobileGLResult that ServerApplyWireSurfaceOp already computes and drops. Without it a client cannot tell a server that died mid-op from a legitimate refusal, and therefore cannot choose between arming the device-lost latch and returning EGL_FALSE.	—	the client's control demultiplexer	hs
SurfaceReply.defaultFb	(deprecated). The encoder passes a literal 0 offset (SurfaceOpCodec.cpp:184) and the decoder never reads it; kEventSurfaceChanged is the landed carrier (P5f fc). The slot is burned, never deleted — a deleted field frees its vtable slot for the next append.	—	nobody, by construction	hs

asio was considered for the byte stream and declined (2026-09-21). It is already vendored and on the include path, so the question is a fair one. But spawn needs five platform mechanisms and asio covers one: its own socketpair returns operation_not_supported on Windows (3rdparty/asio/include/asio/detail/impl/socket_ops.ipp:683), it has no descriptor passing at all, segment delivery is POSIX-only by design, and there is no fork. So adopting it would not make Windows work; it would only put asio headers inside Transport/, which ITransport.h keeps dependency-light on purpose. Revisit at P6.5, where asio::generic::stream_protocol carries a raw (family, type, protocol) triple and therefore gives AF_VSOCK from the same implementation (2026-09-22: the end state's link is TCP, not AF_VSOCK - §10.3 - and the argument is the same for asio::ip::tcp) - and where ITransport lets an AsioTransport land beside SocketTransport without touching a caller.

| LinkTerms{dataPlane, wireForm, maxReplyBytes, cmdWindowBytes, stageWindowBytes} | Appended to Hello/Welcome. dataPlane has exactly one legal value in P6, SharedSegments; wireForm exactly one, StructImage. Anything else is refused by name. The four sizes are stated by the SERVER. | any other value → Refuse{ProtocolMismatch} | the handshake | hs → not landed (verified 2026-09-22: protocol.fbs has no LinkTerms); re-homed to P6.5 nd, §10.3 | | Refuse{code, detail, peer values} | Appended CtrlMsg union tag. A handshake disagreement is an answer, not an abort (§5.1). | — | both sides | hs → not landed (no Refuse in union CtrlMsg); re-homed to P6.5 wf, §10.3 | | wireFingerprint / buildFingerprint | The single abiFingerprint splits. §4. | mismatch → Refuse, never Fatal | the handshake | hs → partially landed (explicit build stamp, abiMajor/abiMinor checked, real pids; the split itself and the PipeFields.def layout digest are not); re-homed to P6.5 wf, §10.3 | | FatalCode | Unchanged, 7 values (protocol.fbs:248-256). table Fatal appends family: string. §5.2. | — | — | dl |

All schema edits are append-only. A FlatBuffers table field's id is its vtable slot and an enum value is the wire value; retired fields become (deprecated), never disappear. CapsSnapshot's four deprecated fields are the worked precedent and their reasoning is in protocol.fbs itself.

hs implementation note, from c6's review. flatc drops deprecated fields from the generated Create* helper. Deprecating defaultFb while appending result therefore changes CreateSurfaceReply's arity, and the existing 6-argument call at SurfaceOpCodec.cpp:183-184 would silently bind its /*defaultFb=*/0 to result. hs converts that call site to SurfaceReplyBuilder with explicit add_* in the same commit, so a positional mistake is impossible rather than merely unlikely.

---

§3 Role, Dial, and the six predicates

3.1 D1 — what Transport reads as in the server process

TransportMode::Spawn. The child does not force Monolith.

ARCHITECTURE.md:488's literal instruction ("mobilegl_server_main hard-sets MG_Config::Transport to Monolith before reaching Init()") is amended by this file. a6 measured why: 226 live lines test Transport != Monolith to select the server arm, 148 of them in MG_Backend. Forcing Monolith flips all of them to frontend glue in the one process with no frontend. The sharpest instance is PipeInputs::IsLive() (MG_Impl/Pipe/PipeFill.cpp:2135): forced monolith, it reads MG_State::pGLContext, which is null there, so 58 MGB_CTX_LIVE guards all answer "not live".

Anti-recursion moves to a separate, non-inherited axis: MG_Config::Dial == DialMode::No in the child. The envp scrub stays as the second, independent catch, and S3 (§9) still falsifies it — and the diagnostic must name which of the two fired, or the double safety is untestable as two things. Package: sm.

3.2 D1c — there are six predicates, not three

a6 proposed three. Three do not cover the tree.

predicate	asks	pull-build shape
MGPipeSplitActive()	do records cross a role boundary at all	#else returns false
MGPipeServerArm()	am I executing as the server right now	#else returns false
MGPipeShouldDial()	should this process connect out	#else returns false
MGPipeSessionLive()	is a peer session live	role-defined; #else false
MGPipeBlocksAreDistinct()	do the two roles share one storage object	#else false
MGPipeBackendIsLocal()	is the backend in this process	#else true

ContextEpoch::Server and ActiveXfbState's role index are MGPipeSplitActive() sites, not MGPipeServerArm() sites: they are partition keys, and the ServerArm disjunct would make a partition key thread-dependent.

D14 — all six must fold away in the pull build, and the required shape is SlotCaps.h's: the non-disaggregated build gets a different definition under #else, not a runtime-false predicate. MGPipeSplitActive() must additionally remain exactly one load of a process global and must never be latched into a static Bool. G1 is 0/0/0/0 and .text unchanged; anything else is a c6 defect. Package: sm.

3.3 D13 — rename before introducing

MGPipeRoleSplitActive() (MG_Backend/MGPipe/PipeInputs.cpp:292) means Ipc.RoleSplitState && Transport != Monolith — the rehearsal knob, default off. It is renamed to MGPipeRoleSplitRehearsalActive(), which is what it has always meant, and MGPipeBlocksAreDistinct() is introduced beside it as the wider predicate. a6's shape — rename, then widen the same function — is overturned: a widened function under a narrow name is a name that lies.

Blast radius, measured at this head: 12 sites, of which DualBlockScenario.cpp's assertion message embeds the old name and must be updated with it. Package: sm.

3.4 D10 — the four guards that cannot arm in a server-only process

guard	today	becomes	package
MGPipeApplierReset layer 2 (MG_Pipe/PipeApply.cpp:1353-1355)	needs ClientSession::Active() != nullptr, permanently null in a server process	middle conjunct → MGPipeSessionLive(); third → !MGPipeServerArm()	sm
MGPipeServerBlockNoteIdentity (PipeInputs.cpp:314)	early-returns unless the rehearsal knob is on	gate → MGPipeBlocksAreDistinct()	sm
CountBarrierPull's unconditional Fatal arm (PipeInputs.cpp:174)	knob-gated	rides MGPipeBlocksAreDistinct()	sm
the two RefuseFromApplyThread helpers (MG_Impl/Pipe/SlotAllocator.cpp:29-57)	ServerLoop::OnApplyThread()	OR'd with a process-role fact set once by ServerMain	sm

The one that crashes rather than merely failing to arm. With MGPipeServerBlockNoteIdentity early-returning, a spawn server's gPipeInputs.ContextIdentity() stays nullptr for the life of the process, and DirectGLES.cpp:199-213's fb-slot memo does a raw pointer compare with no generation against it — so nullptr == nullptr reads as a cache hit and the first call uses an uninitialised slot. PipeInputs.h:878-884's own comment says a named Fatal was intended there. sm lands the gate change; st lands the named Fatal.

---

§4 The handshake (hs)

4.1 One equality test answers three different questions today

AbiFingerprintInputs (Transport/SessionRings.h:544-566) mixes wire properties, struct-layout properties and pure build identity into one Uint64. It splits:

wireFingerprint — always compared. Inputs, in order: an ordered catalogue digest over PipeCalls.def's (opcode, payload struct name, call-flag row, WaitClass); a payload layout digest derived from PipeFields.def (per member: name, offsetof, sizeof); the record-header layout; both blob-codec version stamps; the format-capability table's two extents; kOpCount; MOBILEGL_ABI_VERSION; endianness; pointer width; and sizeof(MGPCaps) and sizeof(DynamicBackendParameters).

Those last two stay compared because they are wire facts, not build facts: MGPWireRec_GetCaps is composed from sizeof(MGPCaps) (MG_Pipe/generated/PipeWire.inc). sizeof(GLFunctionsTable) is deleted from the mix: it never crosses the wire, so adding one backend slot today invalidates every peer for no wire reason.

Why the layout digest is not optional. SessionRings.h:561-565's own comment says the git stamp is in the mix because two builds of the same sizes can still disagree about a FIELD ORDER, which no sizeof can see. A catalogue digest over (opcode, struct name, flags, WaitClass) sees neither a member reorder inside a payload struct nor a width change. So the widely-proposed "drop the git stamp, add a catalogue hash" is wrong as usually stated — it removes a guard and does not replace it. The layout digest is the replacement, and gen_pipe.py already CI-enforces that PipeFields.def names every member, so it is a generator addition and not a toolchain.

buildFingerprint — compared only while Dial == Fork. Under fork the child is the same binary, so comparing it catches a stale MOBILEGL_IPC_SERVER_PATH pointing at an old libMobileGLServer.so — a real P6 failure mode. P6's policy is therefore strictly stronger than today's, not looser; the end state is a policy flip, not a redesign.

4.2 The missing build stamp

a6 found CMakeLists.txt:188-195 takes the git hash with no RESULT_VARIABLE and no fallback, so a failed git rev-parse silently yields "". Verified live: building this worktree from WSL produces GIT_COMMIT_HASH_SHORT "", because the worktree's .git file holds a Windows path WSL's git cannot resolve. Two builds from different commits then get identical fingerprints and the stamp silently stops doing its documented job.

The fix may not touch GIT_COMMIT_HASH_SHORT. That macro is baked into glGetString(GL_VERSION) (MG_Impl/GLImpl/Getter/GL_Getter.cpp:614), so redefining it changes an application-visible GL string and the pull build's .rodata — a behaviour leak and a G1 break at once. Instead MGGitHash.h.in gains a second pair, MOBILEGL_BUILD_STAMP_VALUE and MOBILEGL_BUILD_STAMP_PRESENT, consumed only by the fingerprint, and the CMake step gains a RESULT_VARIABLE. A build with no stamp sets PRESENT to 0, and a Dial == Fork handshake whose peer reports PRESENT == 0 is a named refusal: under fork the two peers are supposed to be the same binary, and without a stamp that premise cannot be checked at all. Package: hs.

4.3 Two verified asymmetries hs also fixes

• The client checks only abiFingerprint and never abiMajor/abiMinor (ClientSession.cpp:566-568).
• Hello::pid is hard-coded 0 and Welcome::serverPid echoes the client's, so it is always 0 (ClientSession.cpp:519-521, ServerSession.cpp:532-534). §9.5's arm-proof gate wants the child pid; serverPid is its natural carrier and must start carrying it.

4.4 Segment geometry (D-size)

MOBILEGL_IPC_RING_MB / STAGE_MB are read from the client process's env with ceilings of 1024 / 4096 MiB (ConfigLoader.cpp:363-364), while the server creates the segments. Under spawn that is a guest environment variable sizing an allocation inside the renderer process. The four sizes become server-stated LinkTerms fields; the client's values become a request the server may clamp, and the clamped values are reported in Welcome.

ShmSegment::Adopt's fstat "do not trust the peer's declared size" check (Transport/ShmSegmentPosix.cpp:127-133) is today the only bound, and it has no analogue on a link with no fd. The compile-time ceiling that refuses an oversized announcement by name is P6.5's to define; P6 does not pretend to have answered it.

---

§5 Death, refusal, and what P5e changed

5.1 The handshake never aborts

A version or terms disagreement is a Refuse{code, detail, peer values} control message plus MOBILEGL_ERR_PROTOCOL_MISMATCH on both sides: the server sends Refuse and exits 0, the client reaps the child and refuses by name, never falling back to monolith. Both FatalAbiMismatch call sites go (ServerSession.cpp:471-483, ClientSession.cpp:565-568). The worked shape is eight lines away in the same function: ServerSession.cpp:456-466 already returns MOBILEGL_ERR_PROTOCOL_MISMATCH for a malformed frame, with the correct msg_as_Hello() != nullptr guard. Package: hs.

5.2 Fatal{} keeps aborting in P6, and gains one funnel

a6 censused 92 abort sites under MG_Remote/ carrying 30 distinct family words, against a 7-value wire FatalCode. So Session::Fail(FatalCode, detail) as usually proposed cannot carry today's vocabulary in its first argument.

D2 ruling. The first argument is an internal MGFatalFamily, generated from a new MG_Remote/FatalFamilies.def — one row per family word, each row carrying its FatalCode projection. The wire keeps FatalCode unchanged; table Fatal appends family: string. The projection is a total table, not a switch with a default: arm — this build has no -Werror anywhere (verified: CMakeLists.txt, every *.cmake, both workflows), so a gate premised on -Wswitch failing the build is fiction.

Session::Fail stays [[noreturn]] in P6 and still ends in abort(). Log lines, CI greps and every red-once are byte-identical. What the funnel buys, and what makes it worth dl's two days:

• one place that publishes a SessionFault{family, detail, seq, op} frame to the peer before dying, so the guest's log names which record killed the session instead of reading a bare EOF;
• one telemetry point;
• a CI census gate refusing any new bare abort outside the funnel.

LANDED, in three steps. (1) FatalFamilies.def + FatalFamily.h generate MGFatalFamily, its total projection FatalCodeForFamily and FatalFamilyName; 32 rows, tested. (2) SessionFail (FatalFunnel.{h,cpp}) is the funnel: the ~90 scattered MGLOG_F("...Fatal{...}"); abort(); pairs under MG_Remote/ become one call each, the message string passed verbatim so every line and grep is byte-identical, the family enum riding alongside for the coarse code. SessionFaultCount() is the telemetry point. The census now also requires every SessionFail/WireLogFatal call's string to carry a Fatal{ word, and its baseline drops from 93 bare aborts to 4 (both funnels). (3) table Fatal gains family: string; SessionFail publishes the frame on whichever session control plane is active before aborting, and the client's control-reply loop recognizes it, names the family and latches device-lost. seq/op are NOT threaded through the 90 sites — the message already names the op — so the frame carries code, family and message, not a field always 0. A kill -9 bypasses abort() and is caught by dl's hangup witness instead; the self-detected fatal is the one that publishes.

The policy flip is not P6's. Making Fail return needs a real return path invented at every site whose callers rely on [[noreturn]], and a6 counted how many do. That is Ph's (§11).

Two sites carry no Fatal{ marker at all (Client/PersistentMapTracker.cpp:1052-1058, Transport/WireLog.cpp:41-50), so "the log stays verbatim" is not true of them and the family grep cannot see them. dl gives each a family tag.

5.3 The device-lost latch does not exist, and P6 writes it

P6-CONTRACT-DRAFT §5.1 says the latch "is implemented unchanged". It is not implemented at all: FatalCode::DeviceLost/ServerCrashed have zero producers and zero consumers, glGetGraphicsResetStatus returns GL_NO_ERROR unconditionally (MG_Impl/GLImpl/Getter/GL_Getter.cpp:2855-2861), and MOBILEGL_IPC_RESPAWN / IDLE_EXIT_S are parsed nowhere (only a comment at Config.h:466). Today a dead server leaves the client silently DECLINE-ing every verb forever with glGetError clean.

It is a package (dl), not carriage. The latch is session-scoped and set from Doorbell::PeerHungUp() — D5c amends "Dead()" here, and §5.4's block says why — never from a timeout. MOBILEGL_IPC_RESPAWN=1 is a named refusal until some stage implements the re-push.

LANDED. ClientSession::DeviceLost() / LatchDeviceLost(), armed at the verb barrier, at the forced run-ahead wait and at the control-reply wait, and read by GetGraphicsResetStatus, which now answers GL_UNKNOWN_CONTEXT_RESET. Not GUILTY or INNOCENT: those assign blame, and the server died for a reason that never crossed the wire — it could as easily have been a driver fault as this client's draw. The consult and its include sit wholly inside MOBILEGL_BUILD_DISAGGREGATED; G1 re-measured after it, 0 symbol diff and .text unchanged at 0xa52203. That measurement also caught an unrelated 16-byte drift the guards did not: P6's O_APPEND log-sink fix was compiled into the pull build too, and is now guarded for the same reason.

5.4 Slow, dead, and frozen are three states

• A published, never-applied record is discarded at the latch, not waited on.
• Slow and dead are distinguished without a timeout threshold. A server one frame behind is P5e's intended steady state; a server that is gone is Fatal{ServerCrashed}. SocketDoorbell already latches m_dead on EOF — take the fact from there.
• A frozen peer produces no hangup, and a6 found the tree has no defence: the control-reply wait m_controlDone.wait(lock, pred) (Server/ServerLoop.cpp:724) has no deadline, while the ring side has kBarrierTimeoutMs. A SIGSTOP'd or freezer'd server wedges the GL thread inside eglMakeCurrent with no diagnostic.

D5 ruling — two waits, two answers. (a) The server's in-process mailbox wait gains a bound and its expiry is Fatal{ControlReplyTimeout, "<op>@<seq>"}. (b) The client's cross-process reply wait is bounded by MOBILEGL_IPC_CONTROL_TIMEOUT_MS (default 5000) and its expiry is not fatal: the client consults the doorbell's death latch — dead latches device-lost, alive-but-silent is a named diagnostic. Packages: cp (a), dl (b). (P7, p7/spawnhang: "silent" is now literal — a server whose apply thread is running the op says so every 250 ms with SurfaceProgress, and each report restarts the client's bound, up to 120 s; a server that never took the op, or is frozen, sends none.)

D5c — the client's self-bell CANNOT witness the server's death, and sm is why. Measured, not predicted: minecraft-1.21.4-fabric-iris-iterationrp-in-world kills the server process outright on this machine (llvmpipe: LLVM ERROR: Cannot select: intrinsic %llvm.x86.vcvtps2ph.256, SIGABRT at call 159429), and the client did not take the SessionWait::ShutDown path that already exists and is already correct. It sat out the full 120 s and reported Fatal{BarrierTimeout, "ResourceRespecify"} — the wrong diagnosis, at 120 s a case, for a server that had been gone since the first second. Under monolith the same trace fails in 3 s.

The cause is structural and is the price of the two-fd bell: the client parks on clientBell[0] (slot 5) and rings itself through clientBell[1] (slot 6), so it holds both ends of that socketpair. EOF arrives only when every writer closes, and the client is one of them — so that descriptor can never hang up, no matter what happens to the peer. A bell that can wake itself is a bell that cannot hear a death.

dl therefore needs a witness descriptor whose far end only the SERVER holds, polled beside the park fd and never read from, whose POLLHUP/EOF latches Dead(). The control socket is already exactly that and needs no new descriptor. The fix does not belong to sm: holding both ends is required for self-notification and is not the defect — assuming one descriptor could answer both questions is.

LANDED. SocketDoorbell::SetDeathWitness(fd); the client hands it SocketTransport::StreamFd(). The witness enters the poll set with events = POLLRDHUP and nothing else, which is what lets it watch a socket it must never read: POLLHUP/POLLERR/ POLLNVAL arrive in revents unrequested, so the hangup is seen while POLLIN never is — a control reply queued on that descriptor can neither wake the bell nor be consumed from under SocketTransport's reassembler. Measured: kill -9 on the server is noticed in 0.01 s against a 120 000 ms barrier, and with the witness removed the same test takes the full 2 s retry budget and fails naming this paragraph.

Dead() AND PeerHungUp() ARE TWO FACTS, and only the second may arm the latch. Park sets m_dead for POLLERR, POLLNVAL and unrecognised revents as well — each a fault in this process's descriptor — and CondVarDoorbell::Kill() sets it on every orderly Stop(). A latch taken from Dead() would report a lost GPU for a local fd bug and a context reset for every clean exit. ADeadBellIsNotAlwaysAHungUpPeer is the red-once for the distinction; AnOrderlyStopIsNotADeviceLoss is not — it stays green under the collapse, because under spawn the self bell is a SocketDoorbell and Stop() never kills one. That was found by running the falsification rather than by reading the code.

---

§6 The control plane (cp)

6.1 What fc landed, and what is left

P5f package fc framed the twelve Server* forwarders as value-only SurfaceControlFrames and landed the wire codec. The draft concluded cp is "transport only". Three things are not transport, and a6 found all three:

1. The InitCapabilities request has no wire form (§2).
2. SurfaceReply cannot carry a MobileGLResult (§2).
3. Neither side of the control plane is wired to a socket. The apply thread's ready predicate has exactly three terms — stop, ControlIsPending(), cmdHead != LocalTail() — and a socket is not among them (ServerLoop.cpp:398-402), so a SurfaceOp arriving on the wire is read by nobody. The client's pump consumes only CapsSnapshot and drops everything else with a WARN (ClientSession.cpp:1425-1432).

6.2 D4 — who reads the control socket

A dedicated mgl-srv-io thread: ReceiveFrame → DecodeWireSurfaceOp → post into the existing one-slot mailbox as an ordinary non-apply-thread poster. ServerLoop.cpp:673-727 is not changed. The io thread is additionally the server's only sender on the control transport: PublishCapsSnapshot enqueues and the io thread drains one outbound FIFO — which is what fixes the ordering hazard that CapsSnapshot is pushed from inside the dispatch, on the apply thread, mid-op (ServerLoop.cpp:959-969).

Rejected: adding a fourth term to the apply thread's ready predicate. That puts frame decode on the thread holding the native context and makes every control arrival a wake of the applier.

6.3 D7 — seq

The client mints, always, dense from 1, never reused in a session. seq == 0 arriving on the wire is Fatal{ProtocolCorruption, "SurfaceOp.seq"} at the decoder, not at RunSurfaceControlFrame, whose local mint (ServerLoop.cpp:659-662) stays and becomes reachable only for a local, non-wire post. Correlation lives in the client's control demultiplexer; DecodeWireSurfaceReply keeps its void signature and stops being the place correlation was expected.

6.4 ForgetCurrentTuple

The N-3 rule holds identically on the frame path: the forgetting calls are inside the dispatch (ApplySurfaceControlFrame) and ServerLoopTest's C7/N-3 controls drive them through framed posts unchanged. No change.

---

§7 Windows — rule H's enforcement (cp)

7.1 The blacklist is open, and the end state's first guest walks through it

WindowBackendForWireWindowKind (Protocol/SurfaceOpCodec.cpp:104-113) accepts WindowKind::X11 → WindowBackend::X11 and Win32Hwnd → Win32. DecodeWireSurfaceOp (:141-179) refuses only AndroidNativeWindow and MetalLayer. An accepted kind's nativeToken is copied through and UnpackWindowHandle (Server/ServerLoop.cpp:875-882) casts it to void* unconditionally, whence it reaches eglCreateWindowSurface in the server process. DetectWindowBackend() is a compile-time #if chain (MG_Impl/EGLImpl/EGLImpl.cpp:59-70), so a guest cannot avoid announcing its own window system.

7.2 D8 — whitelist, and in P6 the accept set for a window op is empty

• X11 → Fatal{UnmigratedSurface, "X11@P12"}; Win32Hwnd → "Win32Hwnd@P12".
• None on a window op → Fatal{ProtocolCorruption, "SurfaceOp.windowKind"}. Today it is accepted, maps to WindowBackend::Unknown, and the raw token is copied through anyway.
• Separately and independently: nativeToken != 0 on any op → Fatal{ProtocolCorruption, "SurfaceOp.nativeToken"}.
• Pbuffer / Surfaceless are surface shapes, not window backends, and remain the only things P6 lands.

Do not append WindowKind::ServerOwned in P6. The schema is append-only, so an enumerator appended now and re-meant later is a wire value burned; P6 has no consumer for it. P12 appends it.

7.3 The client emits unconditionally, and P6 must decide what that costs

Client/BackendObject_Remote.cpp:214-216 posts SetWindowHandle with no split-arm check, so on a real Android app the first eglCreateWindowSurface aborts the server process and the client sees only a dead peer. cp mirrors the refusal on the client side, so the diagnostic names the cause in the process that caused it. This is P6's, not P12's, because P6's device lanes run on a phone.

7.4 m_windowHandle has four stores, not one

a6's precise shape, correcting the corpus (and P6-ENDSTATE-REVIEW §2.5): four stores (MG_Backend/BackendObject.cpp:475, :365, :206, :207), five reads, and :283 is a call, not a store. .Width/.Height have no readers. The smallest latch point for a server-owned window is RegisterEGLWindowSurface (:232-238), and the non-obvious fourth edit is the sameHandle dedup key (BackendObject_DirectGLES.cpp:962-967), which under server ownership never matches and would rebuild the native context on every eglCreateWindowSurface. Recorded here for P12; P6 changes none of it.

---

§8 The data plane (lk implements, c6 declares)

8.1 Why a seam at all

ITransport.h:16-20 says in its own header that the hot path bypasses it: it is the control plane, and that scope is already correct for a socket. Below it there is no interface of any kind — Ring.h, ReplySlot.h and EventRing.h take raw void* base + capacity. So "swap the transport" has nothing to swap.

c6 declares Transport/ILink.h and Transport/StreamLink.h. lk implements ShmLink by moving today's SessionSegments / Ring / ReplySlotPool / EventRing behind the interface, line for line, with zero behaviour change.

8.2 The seam is a setup-time interface, not a hot-path one

ILink hands out POD handles once at attach and then stays out of the way. No ILink virtual is called per record, per publish, or inside any spin predicate. Three grounds, and the second is decisive:

1. Two virtuals per record at the 4000 records/frame that SEG_CMD is sized against is ~8000 indirect calls/frame — about 0.5% of the measured 15.72 ms client frame. Affordable, but the only lane that could measure it has ±4% run-to-run dispersion, eight times the effect, so the gate P6-ENDSTATE-REVIEW demands could only ever print "within noise" — a green that proves nothing.
2. The spin predicate forbids it outright. Doorbell::Wait is a non-virtual template whose spin loop calls ready() up to SpinItersPerUs() * spinUs times; at 916 waits/frame with 99.8% resolved inside the spin, the predicate runs 10⁵–10⁶ times per frame and reads the watermarks straight off the page. That is the 31.58%-of-the-GL-thread path. A virtual there is not a rounding question.
3. The encode path already depends on four inline accessors around its single Reserve, and virtualising Reserve alone would force either four more virtuals or moving R-10's wrap counters behind the seam — and those are deliberately encoder-owned.

The gate is therefore an objdump call-count of the encode path, not a wall clock: after lk the path must have no more out-of-line calls than today. That is falsifiable; "within noise" is not.

8.3 RingControl's regroup is lk's, and it is a wire break the day a second build exists

The watermark cache line is {appliedSeq, submittedSeq, retiredSeq, completedFrameSerial, presentAckSerial}; eventRingFull/eventDropped sit in a separate alignas(64) group with serverEpoch/ringGeneration/the two park flags. And submittedSeq is producer-written (Ring.h's "THE FIVE WATERMARKS" block) while the other four are consumer-written — so the group has mixed writers, and any "single-writer progress block aliased over the watermark line" is unsound.

lk therefore regroups: submittedSeq moves to the producer's own line, beside cmdHead, which is the other thing the producer writes; what is left — the four consumer-written watermarks — becomes a named LinkProgress member with per-field offsetof static_asserts.

Correction, found while lk was measuring its own blast radius. An earlier revision of this section, and of ILink.h, also promoted eventRingFull/eventDropped into the watermark group and called the result single-writer. Both were wrong. eventRingFull has two writers: the server latches it with store(1) when SEG_EVENT fills (Transport/EventRing.h:140) and the client clears it with an exchange(0) on every drain (:200). Promoting it would put a per-drain RMW by the client on the same cache line as appliedSeq — which the server release-stores every 64 records and the client's spin predicate reads 10⁵–10⁶ times per frame. That is new false sharing on the hottest line in the system, and it would land in exactly the numbers this package's gate has to hold constant. The two flags stay in the doorbell/generation group, where mixed writers already live beside consumerParked/producerParked, and cross the seam through a separate EventFlags() accessor. Accessor, not line: they share serverEpoch's line, which is where mixed writers already live.

It is free today — both peers are the same binary, enforced by the fingerprint — and a wire break the day a second build exists. It is lk's and not c6's because the rename blast radius is 109 sites in production alone, not the 39 an earlier draft assumed, and because five of them sit behind #if defined(MGITEST_SPLIT_RUNTIME_PEEK) && !defined(__ANDROID__) — an Android configure compiles them out, so a missed rename is green on device and red only on a desktop split-peek build. lk publishes the per-file itemisation before it starts.

8.4 What StreamLink is for

Every method is a named refusal, Fatal{UnmigratedDataPlane, "StreamLink@P6.5"}. It exists so the seam's adequacy is a compile, not an argument: if ILink cannot name what a byte stream needs, StreamLink cannot be declared against it. Writing it is what revealed that the seam needs FlushProgress(), ResolveSpan() and a reply pair — three methods the brief's own list (progress transmission, send window, flush) did not name.

c6's red-once: delete one ILink method StreamLink declares and the build goes red.

8.5 The purity gate (lk)

After lk, the identifiers RingControl, RingProducer, RingConsumer, ReplySlotPool and EventRingConsumer, and any m_shm. expression, appear only inside Transport/ShmLink.* and the transport's own tests. Same shape as the existing A/B/C purity gates. The baseline is recomputed mechanically at lk's head and published per file — the earlier draft's ratchet number was wrong by a factor of three, and a ratchet installed at the wrong number is red on day one for reasons that have nothing to do with a leak.

8.6 SEG_STAGE's cursors are deliberately dead

Ring.h's "DEAD IN P5, DELIBERATELY" states it and a test pins it: SEG_STAGE is not a ring; staging is an encoder-local linear allocator reclaiming on retiredSeq, and all three stage cursors stay zero for the whole of P5. A stream link's send window either revives them completely or does not use them; the header names the middle state as a guaranteed hang rather than a slow path. P6.5's, recorded here so it is not rediscovered.

---

§9 The gate

Beyond the five-part gate (docs/Disaggregated/ARCHITECTURE.md §13.2):

1. G1 — pull build 0/0/0/0, .text unchanged. Every predicate in §3.2 folds away.
2. G2/G14 — integration-spawn's case-name set identical to integration-split's.
3. integration-spawn green at the same count as integration-split.
4. Process tree, mechanical — exactly one extra child while running, zero after; HeadlessGL's fork pre-check leaves no orphan.
5. Arm proof (ID-124) — every spawn entry records the child pid and transport=spawn in its own private log. A spawn lane that fell back to monolith passes items 1–4. §1 is why this gate is not optional: today the fallback and the intended state are the same value.
6. Red-once per package (R-16), one named pair each.
7. Device — Redmi 2f7cbe2e, reboot-clean, paired A/B in one thermal window.
8. Performance recorded, not gated, but three numbers are mandatory: the socket doorbell's cost against inproc's condvar; SEG_STAGE bytes/frame; and records/frame post-chunking. The last two are takeable on the existing inproc lane today and MEASUREMENTS.md:342 says in as many words that no post-chunking distribution exists.

Item 8's wording must be able to accept a negative regression without looking suspicious: today the client spins ~1600 times/frame on a shared cache line with 99.8% of waits resolved there, and a socket link has no such line. Converting that into ~4 blocking reads/frame may be cheaper.

Negative controls, each run red once:

#	knob	what it must falsify	package
S1	MOBILEGL_IPC_SERVER_PATH=/nonexistent	named refusal, not a silent monolith fallback	sm
S2	kill -9 the server mid-frame	the latch; the client never hangs and never keeps submitting. Red-once the wrong implementation too: set the latch from a timeout and the paired control "a server one frame behind is the normal state" must go red	dl
S3	child envp not scrubbed	Dial = No catches it and the diagnostic names which of the two catches fired	sm
S4	inherited P5f epoch invalidation disabled	a replaced context inherits stale synced state; a still-live XFB lifetime retains its paused span	st
S5	WindowKind::X11 with a plausible XID reaches the server	named Fatal, and the gate falsifies that the token reached UnpackWindowHandle. Today's S5 tests only AndroidNativeWindow and stays green with §7.1's hole open	cp
S6	a peer offers LinkTerms.dataPlane = StreamOnly	refused by name, and the server survives to serve the next connection	hs
S7	a bare abort() added outside Session::Fail	the census gate goes red	dl
S8	—	integration-spawn asserts SessionFaultCount() == 0 over a whole run: the good-path assertion E1/ID-122 wanted and could not express, and which S5/S6 can falsify	t6

---

§10 Amendments

10.1 To P6-CONTRACT-DRAFT.md, which this file supersedes

1. §1's "ABI fingerprint | Unchanged … same-machine, same-binary" → §4.
2. §5.1's "The latch … is implemented unchanged" → false; §5.3.
3. §4's "P6's remainder is TRANSPORT ONLY" → false in three ways; §6.1.
4. §7's "a transport swap and nothing else" → no longer true. This file adds lk, hs and dl to the package list and says so here rather than smuggling it.

10.2 To the earlier contracts

1. CONTRACT-P5C.md:537-539 — EGL value framing was discharged by P5f fc, the unpack/render-state epoch work by P5f fs. P6 carries the existing frames over the socket and verifies the inherited lifecycle rules in the second process.
2. CONTRACT-P5C.md:56 — P6's multi-context shape. Declined: P6 stays single-context. §11.
3. Transport/ReplySlot.h's "P6 debt" for the 2 MiB reply cap and chunked readback. Declined: P9's. Note P6-SPAWN-PLAN §8.5 already reassigned it and the header was never updated — two documents in the tree disagreed; this file settles it.
4. Ring.h's "BATCHING MAY ONLY MAKE A WATERMARK LATE" — appliedSeq is excluded from lazy publication by name. P6 does not change the behaviour; it records that any future stream link amends this rule rather than inferring around it. Several designs have argued the late-never-early invariant licenses lazy appliedSeq; it explicitly carves appliedSeq out.
5. ARCHITECTURE.md:488 — the child-forces-monolith rule. Amended; §3.1.
6. ARCHITECTURE.md:9 and MG_Backend/MGPipe/PipeInputs.h:281, :744, :957, and P5F-WIRE-COMPLETENESS §1.4 — "the server does not link MG_Impl". False at this head. §12.

10.3 Amendments of 2026-09-22 — the end state re-ruled

1. The end state is TCP across machines, OSes and architectures, not an AVF pVM over AF_VSOCK. docs/Disaggregated/notes/p6/P6-ENDSTATE-REVIEW.md carries the retraction at its head; its §3 gains form D. Same-machine spawn (AF_UNIX + shared segments) stays as the local form.
2. The transport stack has two independent axes. The control plane (ITransport: fork inherited fds / unix:<path> / tcp://host:port) and the data plane (ILink: ShmLink / StreamLink) are chosen separately, negotiated in the handshake, and may be mixed — on one machine, control over TCP with data over shared segments is a legal pair. Callers above Transport/ see ITransport + ILink and nothing else; branching on the link kind above that line is a purity-gate failure (the lk grep gate, widened). Design: ARCHITECTURE.md §11.9.
3. §8.2's "ITransport is not extended" is amended in one respect: descriptor passing leaves the control plane and becomes the shared-segment data plane's delivery. ShmLink brings its own AF_UNIX aux rendezvous, named in Welcome, and SCM_RIGHTS travels on that and nowhere else. Rule G is unchanged: a transport that declares no descriptor passing (TCP) passes none, and the data plane that needs it must bring its own channel or not be selected. auto selection is decided by a successful delivery, never by inspecting addresses, and a fall-back to Stream is named in the log and the stats line.
4. §2's three hs rows (LinkTerms, Refuse, the fingerprint split) did not land in hs; protocol.fbs carries neither LinkTerms nor a Refuse member of CtrlMsg. They are P6.5's (nd, wf), and S6 in §9 moves with them. CURRENT_STAGE_PROGRESS.md's earlier claim that S6 had been run red is withdrawn.
5. Fixed-width wire form is no longer negotiable-later. Two different binaries from two compilers are the end state, so the PipeFields.def-derived layout digest, the fixed-width rewrite of DynamicBackendParameters / MGPCaps / the RenderStateParameters blob, and the buildFingerprint-only-under-Dial == Fork policy are P6.5 wf, taken off P7.
6. Ph gains pairing/authentication: a Hello without a pairing token is refused on any non-loopback listener. Ph precedes P12's non-loopback listen.
7. "The 511 class-C entries" in §11 is the P5-joint census (docs/Disaggregated/notes/p6/census-classC.md). At this head the MGR_UNMIGRATED_* lists are empty; SetSwapInterval (P10) and DeleteTransformFeedback (P9) are the two that remain.

---

§11 Not P6

Real window arrival and the android:process=":mgl" Service (P12); multi-context (P12); chunked readback and the reply-slot pool (P9); DynamicBackendParameters' fixed-width rewrite (P7); MOBILEGL_IPC_POLL_ESCALATE (P10); remaining monolith-only frontend-object / twin-registry glue (P3b/P4b) and the named P7 functionality debts; Windows pipe: / unix:; the 511 class-C entries.

The two-axis transport stack — control × data, TCP-capable, cross-build — is P6.5 and is the IPC track's next stage (§10.3). P9 builds its reply semantics on P6.5's message replies rather than on a slot pool, and P11's T0/T1 are refused by name on a stream data plane. Untrusted-guest hardening is Ph, and it must land before any shipping server app accepts a connection from a guest it did not produce — including the Fatal-policy flip (§5.2), the handle-slot budget, and the caps in §12.

E1's re-specification (ID-122) is not discharged by P6 and does not become P6's because P6 starts.

---

§12 Recorded debt

12.1 The server image links MG_Impl in P6 (D12)

P6's server image is the whole libMobileGL.so — which is what ARCHITECTURE.md:496's "one shared library, two roles" always meant. The assertions that the server does not link MG_Impl are false and become debt with named owners.

Measured: 184 symbols, 177 of them real function calls, that the server set needs and only MG_Impl / MG_State / MG_Remote/Client define. Owners: P7 101, P3b/P4b 14, both backends 60, P6's own 6. There is additionally no module target to link — SOURCE_FILES is one flat list feeding both library targets — so establishing module boundaries is itself named work, owed to P13. Method and full symbol list: a6-link-experiment.md.

P6's own six are not P7's and are sm/lk's to clear: Server/PipeApplier.cpp calls Client::ClientSession::NoteApplyThreadEnteredApplier() / NoteApplyThreadLeftApplier(); Wire/PipeWireCodec.cpp needs MG_State::GLState::DecodeProgramArchive and Client::AdoptTierIsEmulate(); MG_Pipe/PipeApply.cpp needs Client::ClientSession::Active() and AdoptTierIsEmulate().

Gate (landed, P7 wave 0 — ID-P7-7): scripts/link_ratchet.py recomputes this set from an explicit per-object partition on every build-linux-split run and fails on any symbol absent from scripts/data/link_ratchet_baseline.txt — which stands at 186 on e8b2c4bd, the 184 plus P6.5's Client::ClientSession::StartSpawned() and MG_State::GLState::ProgramArtifactsSchemaFingerprint(), with the three backend buckets reproducing a6's 101 / 14 / 60 exactly — while a symbol that disappears is reported rather than failed, so progress never turns the build red (link-ratchet.md).

12.2 D11 — five unbounded peer-driven allocations, capped, with no try/catch

Every peer-declared count is bounded before it reaches an allocator and the bound is a named refusal. MG_Remote gets no try/catch: a cap that is checked is better than an exception that is caught, and the funnel (§5.2) is where the diagnostic lives.

site	today	package
MG_Pipe/PipeApply.cpp:1563-1569	resizes to a peer-supplied Cso.Slot + 1 with no slot limit; every sibling create-family goes through RecordAt's bounded path. Slot = 0xFFFFFFFE asks for ~1.7 TB	dl
Server/PipeApplier.cpp:416-441	readback scratch resized from Box.W * Box.H * bpp before any size gate. 65535² RGBA8 asks for ~17 GB	dl
MG_State/.../ProgramArtifactsCodec.cpp:166-175, :216-223	count bounded by remaining bytes, then multiplied by sizeof(element)	dl
Server/StagedTextureStore.h	resize not bounded against the level's declared extent	Ph
SEG_EVENT overflow	a guest that merely stops draining kills the host renderer in 60 s	Ph

12.3 One guest per server process (D9)

For the whole of P6, one server process serves exactly one client session, and it is a structural fact, not a preference: g_Display/g_Context/g_Surface/g_Config are per-process file-statics and InitDisplayAndContext opens with an unconditional DestroyEGLContext() whose last act is eglTerminate — a process-wide EGL teardown. ServerSession::Accept's same-object arm (ServerSession.cpp:425) returns MOBILEGL_ERR_INVALID_ARGUMENT with no log line at all; sm gives it one. Under Dial == Fork a second guest is unreachable by construction, so the invariant becomes falsifiable only under Dial == Connect, which P6 does not land.

12.4 Settled by ruling rather than by experiment

• Who runs the applier in the server process: ServerMain must drive it through ServerLoop::Start(), so g_applyThreadKey is published and the dozen OnApplyThread() guards keep meaning what they mean. Owner sm; red-once drives the applier without Start().
• The server's exit primitive: _exit, after an explicit log flush. It decides whether g_processTeardown is set and therefore whether seven twin destructors take their early-out. Owner sm.
• The handshake asserts the resolved subsystem mask; a mismatch is Refuse, not Fatal. The envp scrub list does not include MOBILEGL_PIPE_*, so the two processes can resolve different masks from their own environments. Owner hs.
• P6 maps no segment read-only. ShmSegment.h:62-63's stated design — the peer gets SEG_CMD/SEG_STAGE read-only on the server side — cannot be implemented as written, and the comment is corrected rather than left to mislead. Owner lk.
• DualBlockScenario's serverIdentity != nullptr is an in-process fact, and G2/G14 requires the spawn lane to keep the case name. Its assertions are re-homed, not weakened: DualBlockPeek becomes a control-plane query answered by the server process. The declared fallback — assert MGPipeBlocksAreDistinct() on the client side and carry the identity half by §3.4's server-side Fatal — must be declared here if taken, not discovered in review. Owner t6.