doc: Add design notes for cluster mempool and explain new mempool limits

diff --git a/doc/policy/README.md b/doc/policy/README.md
index 4392ffb6..a03edc2f 100644
--- a/doc/policy/README.md
+++ b/doc/policy/README.md
@@ -9,7 +9,7 @@ contents. Policy is *not* applied to transactions in blocks.
 
 This documentation is not an exhaustive list of all policy rules.
 
-- [Mempool Limits](mempool-limits.md)
+- [Mempool Design and Limits](mempool-design.md)
 - [Mempool Replacements](mempool-replacements.md)
 - [Packages](packages.md)
 
diff --git a/doc/policy/mempool-design.md b/doc/policy/mempool-design.md
new file mode 100644
index 00000000..52ebd1bc
--- /dev/null
+++ b/doc/policy/mempool-design.md
@@ -0,0 +1,104 @@
+# Mempool design and limits
+
+## Definitions
+
+We view the unconfirmed transactions in the mempool as a directed graph,
+with an edge from transaction B to transaction A if B spends an output created
+by A (i.e., B is a **child** of A, and A is a **parent** of B).
+
+A transaction's **ancestors** include, recursively, its parents, the parents of
+its parents, etc.  A transaction's **descendants** include, recursively, its
+children, the children of its children, etc.
+
+A **cluster** is a connected component of the graph, i.e., a set of
+transactions where each transaction is reachable from any other transaction in
+the set by following edges in either direction. The cluster corresponding to a
+given transaction consists of that transaction, its ancestors and descendants,
+and the ancestors and descendants of those transactions, and so on.
+
+Each cluster is **linearized**, or sorted, in a topologically valid order (i.e.,
+no transaction appears before any of its ancestors). Our goal is to construct a
+linearization where the highest feerate subset of a cluster appears first,
+followed by the next highest feerate subset of the remaining transactions, and
+so on[1]. We call these subsets **chunks**, and the chunks of a linearization
+have the property that they are always in monotonically decreasing feerate
+order.
+
+Given two or more linearized clusters, we can construct a linearization of the
+union by simply merge sorting the chunks of each cluster by feerate.
+
+For any set of linearized clusters, then, we can define the **feerate diagram**
+of the set by plotting the cumulative fee (y-axis) against the cumulative size
+(x-axis) as we progress from chunk to chunk. Given two linearizations for the
+same set of transactions, we can compare their feerate diagrams by
+comparing their cumulative fees at each size value. Two diagrams may be
+**incomparable** if neither contains the other (i.e., there exist size values at
+which each one has a greater cumulative fee than the other). Or, they may be
+**equivalent** if they have identical cumulative fees at every size value; or
+one may be **strictly better** than the other if they are comparable and there
+exists at least one size value for which the cumulative fee is strictly higher
+in one of them.
+
+For more background and rationale, see [2] and [3] below.
+
+## Mining/eviction
+
+As described above, the linearization of each cluster gives us a linearization
+of the entire mempool. We use this ordering for both block building and
+eviction, by selecting chunks at the front of the linearization when
+constructing a block template, and by evicting chunks from the back of the
+linearization when we need to free up space in the mempool.
+
+## Replace-by-fee
+
+Prior to the cluster mempool implementation, it was possible for replacements
+to be prevented even if they would make the mempool more profitable for miners,
+and it was possible for replacements to be permitted even if the newly accepted
+transaction was less desirable to miners than the transactions it was
+replacing. With the ability to construct linearizations of the mempool, we're
+now able to compare the feerate diagram of the mempool before and after a
+proposed replacement, and only accept the replacement if it makes the feerate
+diagram strictly better.
+
+In simple cases, the intuition is that a replacement should have a higher
+feerate and fee than the transaction(s) it replaces. But for more complex cases
+(where some transactions may have unconfirmed parents), there may not be a
+simple way to describe the fee that is needed to successfully replace a set of
+transactions, other than to say that the overall feerate diagram of the
+resulting mempool must improve somewhere and not be worse anywhere.
+
+## Mempool limits
+
+### Motivation
+
+Selecting chunks in decreasing feerate order when building a block template
+will be close to optimal when the maximum size of any chunk is small compared
+to the block size. And for mempool eviction, we don't wish to evict too much of
+the mempool at once when a single (potentially small) transaction arrives that
+takes us over our mempool size limit. For both of these reasons, it's desirable
+to limit the maximum size of a cluster and thereby limit the maximum size of
+any chunk (as a cluster may consist entirely of one chunk).
+
+The computation required to linearize a transaction grows (in polynomial time)
+with the number of transactions in a cluster, so limiting the number of
+transactions in a cluster is necessary to ensure that we're able to find good
+(ideally, optimal) linearizations in a reasonable amount of time.
+
+### Limits
+
+Transactions submitted to the mempool must not result in clusters that would
+exceed the cluster limits (64 transactions and 101 kvB total per cluster).
+
+## References/Notes
+[1] This is an instance of the maximal-ratio closure problem, which is closely
+related to the maximal-weight closure problem, as found in the field of mineral
+extraction for open pit mining.
+
+[2] See
+https://delvingbitcoin.org/t/an-overview-of-the-cluster-mempool-proposal/393
+for a high level overview of the cluster mempool implementation (PR#33629,
+since v31.0) and its design rationale.
+
+[3] See https://delvingbitcoin.org/t/mempool-incentive-compatibility/553 for an
+explanation of why and how we use feerate diagrams for mining, eviction, and
+evaluating transaction replacements.
diff --git a/doc/policy/mempool-limits.md b/doc/policy/mempool-limits.md
deleted file mode 100644
index 73ab017f..00000000
--- a/doc/policy/mempool-limits.md
+++ /dev/null
@@ -1,65 +0,0 @@
-# Mempool Limits
-
-## Definitions
-
-Given any two transactions Tx0 and Tx1 where Tx1 spends an output of Tx0,
-Tx0 is a *parent* of Tx1 and Tx1 is a *child* of Tx0.
-
-A transaction's *ancestors* include, recursively, its parents, the parents of its parents, etc.
-A transaction's *descendants* include, recursively, its children, the children of its children, etc.
-
-A mempool entry's *ancestor count* is the total number of in-mempool (unconfirmed) transactions in
-its ancestor set, including itself.
-A mempool entry's *descendant count* is the total number of in-mempool (unconfirmed) transactions in
-its descendant set, including itself.
-
-A mempool entry's *ancestor size* is the aggregated virtual size of in-mempool (unconfirmed)
-transactions in its ancestor set, including itself.
-A mempool entry's *descendant size* is the aggregated virtual size of in-mempool (unconfirmed)
-transactions in its descendant set, including itself.
-
-Transactions submitted to the mempool must not exceed the ancestor and descendant limits (aka
-mempool *package limits*) set by the node (see `-limitancestorcount`, `-limitancestorsize`,
-`-limitdescendantcount`, `-limitdescendantsize`).
-
-## Exemptions
-
-### CPFP Carve Out
-
-**CPFP Carve Out** if a transaction candidate for submission to the
-mempool would cause some mempool entry to exceed its descendant limits, an exemption is made if all
-of the following conditions are met:
-
-1. The candidate transaction is no more than 10,000 virtual bytes.
-
-2. The candidate transaction has an ancestor count of 2 (itself and exactly 1 ancestor).
-
-3. The in-mempool transaction's descendant count, including the candidate transaction, would only
-   exceed the limit by 1.
-
-*Rationale*: this rule was introduced to prevent pinning by domination of a transaction's descendant
-limits in two-party contract protocols such as LN.  Also see the [mailing list
-post](https://lists.linuxfoundation.org/pipermail/bitcoin-dev/2018-November/016518.html).
-
-This rule was introduced in [PR #15681](https://github.com/bitcoin/bitcoin/pull/15681).
-
-### Single-Conflict RBF Carve Out
-
-When a candidate transaction for submission to the mempool would replace mempool entries, it may
-also decrease the descendant count of other mempool entries. Since ancestor/descendant limits are
-calculated prior to removing the would-be-replaced transactions, they may be overestimated.
-
-An exemption is given for a candidate transaction that would replace mempool transactions and meets
-all of the following conditions:
-
-1. The candidate transaction has exactly 1 directly conflicting transaction.
-
-2. The candidate transaction does not spend any unconfirmed inputs that are not also spent by the
-   directly conflicting transaction.
-
-The following discounts are given to account for the would-be-replaced transaction(s):
-
-1. The descendant count limit is temporarily increased by 1.
-
-2. The descendant size limit temporarily is increased by the virtual size of the to-be-replaced
-   directly conflicting transaction.