docs: FABRIC-4 §3, information state over information moment (Captain Bob's theory)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

docs/fabric/FABRIC-4.md

@@ -104,3 +104,43 @@ addressable VM other patrons talk to, rather than a bare C primitive anyone can
 - How this interacts with §IV.1's own still-open native-boot-flow work — a GPIO VM needs GPIO
   register addresses from the DTB the same way the rest of that bring-up does, so this idea
   isn't independent of that work, just downstream of it.
+
+## 3. Information state over information moment (2026-10-16)
+
+Raised in conversation with Captain Bob during MESH step 7 (`docs/v4.0.0/MESH.md` §9a) --
+theory stage, no code, no scope, deliberately not punch-listed yet.
+
+**The idea, in his words:** "I actually have a theory about what information actually is; the
+temporal existence and not a moment." And: "What we have is a 'information state' over
+information moment. The question I will answer is that is there any advantage to this over a
+conventional OS." Said after "simple is the object."
+
+**Where it came from.** Step 7 first handed a believed GONE on from inside the node that believed
+it, at that instant, on the stacks of whatever text it was in -- information as a moment. The final
+review found it could be lost at depth, and a node between texts could lose its values to a fault.
+The fix that was ruled (2026-10-16) makes the GONE a state: the kernel puts it on each node's
+wire, where it waits until the node takes it. The moment version needed depth sweeps and stack
+constants; the state version needs neither. The same shape runs through the design already:
+messages that wait on wires, not in nodes (7d); patrons that arrive and depart over time (the
+Stadium, `FABRIC-0.md`); K conserved, not sampled.
+
+**The question he will answer:** is there any advantage to this over a conventional OS?
+
+**Contrasts noted in the conversation, as starting points, not conclusions:**
+- A conventional OS delivers much information as moments: Unix signals run a handler at whatever
+  instant they land, on the interrupted code's stack, may call only async-signal-safe functions,
+  and plain signals arriving together coalesce -- the failure step 7's hand-on had. Interrupts,
+  preemption points and callbacks share the shape.
+- What state may buy: no reentrancy class of bug (nothing runs on borrowed stack); nothing silently
+  lost to timing (a fact persists until taken, or is refused at once and its sender told, 7d.1);
+  reproducibility (state can be hashed and compared across machines, as `dict_hash` and K are);
+  verifiability (state is what Isabelle reasons about; interleavings of moments are what make
+  conventional kernels hard to prove).
+- What it costs: latency (a fact waits until it is taken), memory for the queues, and a policy for
+  a full wire (today: refusal, or for the kernel's GONE, not told).
+- How it could be answered with evidence: the DoE apparatus -- the same workload with facts
+  delivered as state and as interrupt-style moments, measuring lost events, variance (CV), and
+  behaviour under pressure.
+
+**Explicitly not decided:** what "information" means formally in this theory; whether it bears on
+the patent work; any experiment design. Captain Bob's question to answer, not yet scoped.