@serfab/quereus-plugin-sereus 0.12.0 → 1.0.0

package/dist/cluster-size.d.ts

@@ -1,3 +1,4 @@
+import type { NodeOptions } from '@optimystic/db-p2p';
 /**
  * Smallest cluster size Optimystic will honour — its own `minAbsoluteClusterSize`,
  * and the smallest value that reaches the cluster path at all (a lone node writes to
@@ -7,7 +8,12 @@ export declare const MIN_CLUSTER_SIZE = 2;
 /**
  * How many nodes each block of the **control** (cadre membership) database is replicated
  * to. Chosen to exceed any party's node count, so in practice every control block lands
- * on every member of the party.
+ * on every member of the party. Cohort width alone only delivers that for blocks written
+ * while the members were serving the network: a block committed while the writer was
+ * alone (the founder's genesis above all — collection headers are written once and never
+ * revised) has a cohort of one, and it reaches later joiners through cadre-core's
+ * peer-join block catch-up (`peer-join-backfill.ts`, membership-gated on the control
+ * network), not through this breadth.
  *
  * **Why full replication here.** Every control node reads the whole control database —
  * membership, peer addresses, the strand list — so a member left out of a block's cohort is
@@ -90,6 +96,20 @@ export declare const CONTROL_REPLICATION_BREADTH = 16;
  * is what the corroboration floor is asking. See Optimystic's `cluster/cluster-policy.ts`
  * ("Why two size yardsticks, not one") and `corroboratorCapacity` in `cluster/quorum-restore.ts`.
  *
+ * **`assumedClusterSize` is now the ADMISSION yardstick.** Since Optimystic grew
+ * `repairCorroborationClusterSize` (see {@link controlClusterPolicy}), the repair corroboration
+ * floor prefers that separate declaration and reads this field only when nothing was declared —
+ * the unknown-machine-count path, where the paragraph above still describes what it does. So on
+ * every path where the count IS known, the 2 here governs one consumer: the membership admission
+ * gate's low-confidence fallback path. Leaving
+ * it at 2 is deliberate and must NOT be "fixed" to match the repair yardstick. Raising it would
+ * make the gate demand `ceil(0.75 x N)` DECLARED peers before it admits a coordinator's peer
+ * set on that path, and a Cadre party is phones and laptops: it cannot promise that three
+ * quarters of its enrolled machines are awake at any moment, so a party of eight with three
+ * awake would stop transacting entirely. The repair yardstick has no such cost (it gates only
+ * whether this node believes a repair answer), which is exactly why the two are now separate
+ * numbers rather than one.
+ *
  * The `satisfies` is load-bearing: without it a mistyped or obsolete key would compile at both
  * this definition and every consumer, because TypeScript only excess-property-checks fresh
  * object literals, not a shared constant handed to `clusterPolicy`.
@@ -125,6 +145,13 @@ export declare const CONTROL_CLUSTER_POLICY: Readonly<{
  * phones and laptops is the ordinary case, not the rare one. Six buys no more fault tolerance
  * than four and costs more overfetch (see the NOTE below).
  *
+ * **Exercised at four machines** by `integration-tests`'
+ * `scenarios/strand-two-party-two-machine.integration.ts`, the only scenario that runs a
+ * strand at this breadth. It shows a write committing while one of the four machines is off.
+ * It does NOT distinguish that from a commit on a cohort downsized to the three live holders
+ * (`allowDownsize` in {@link STRAND_CLUSTER_POLICY}), so the table row above remains reasoned
+ * arithmetic with an end-to-end existence proof under it, not a measurement of cohort width.
+ *
  * **It is also a correctness floor, not only a durability one — but the floor is measured in
  * machines, not in this number.** A node that has fallen behind and can ask exactly one peer
  * whether it is current has Optimystic accept that single answer as the cluster's truth
@@ -133,12 +160,24 @@ export declare const CONTROL_CLUSTER_POLICY: Readonly<{
  * honestly reports the stale revision, the reader concludes it is current, and re-arms its
  * repair window forever. Measured on the control-DB replication scenario: 4 failures in 10 runs
  * at breadth 2, 0 in 20 at breadth 3, 0 in 10 at breadth 8. What lifts the floor off a single
- * voter is a *third machine*, not a wider target: `corroboratorCapacity` takes the cohort peers
- * actually present against {@link STRAND_CLUSTER_POLICY}'s declared `assumedClusterSize`, so a
- * three-machine strand needs two corroborators whatever this number says, and a two-machine one
- * cannot supply them however wide the target is. The underlying Optimystic behaviour is unfixed
- * — `backlog/debt-read-repair-single-voter-corroboration` — so a two-machine strand still takes
- * the exposure. Raising this number does not buy the way out; adding a machine does.
+ * voter is a third machine **or a declaration that there are three** — not a wider target.
+ * `corroboratorCapacity` takes the max of the cohort peers actually present and the declared
+ * repair yardstick, so raising *this* number (the replication breadth) does nothing: the
+ * yardstick is a separate declaration, and **a strand node makes no such declaration today**.
+ * The control network derives one from the party's enrolled machines
+ * ({@link controlClusterPolicy}), which is exact there because every enrolled machine runs the
+ * control node. A strand has no equivalent count: it launches only on machines whose embedder
+ * registered its sApp config, and reusing the party count over-declares — which is worse than
+ * declaring nothing, because at a yardstick of three or more the corroboration floor is pinned
+ * at two peers and a strand that can field only one could then never repair at all. So every
+ * strand runs {@link STRAND_CLUSTER_POLICY} unchanged and takes the single-voter exposure
+ * whenever its cohort view shrinks to one peer. The underlying Optimystic behaviour is unfixed
+ * (`backlog/debt-read-repair-single-voter-corroboration`) and the per-strand serving count that
+ * would let a strand declare honestly is `backlog/feat-strand-yardstick-from-serving-machines`
+ * — its seam ({@link strandClusterPolicy}'s `servingMachines`, threaded through
+ * `StartStrandConfig.servingMachines`) is already in place and fed by nothing. Raising this
+ * number does not buy the way out; a machine, or an honest declaration of the machines actually
+ * serving the strand, does.
  *
  * **Why not derived from the party or member count.** The strand's `Member` rows live *in* the
  * strand database, which runs on the strand libp2p node, whose cluster size is frozen at
@@ -156,11 +195,14 @@ export declare const CONTROL_CLUSTER_POLICY: Readonly<{
  * cadre-driven path, the one that resolves a seed for itself, a strand's libp2p mesh is seeded
  * from exactly one party's own control-database peer rows (`CadreNode.resolveCohortSeed`), so
  * the cohort actually formed is that one party's machines. Four copies are, on that path, four
- * machines belonging to one party — one failure domain, one witness. A cross-party mesh is not
- * forbidden, only undiscoverable: it takes hand-supplied addresses
- * (`StrandConnectionOptions.bootstrapNodes`) or a hand-written dial (see the doc). Tracked:
- * `backlog/feat-strand-party-identity`. Full reasoning: `docs/architecture.md` →
- * "Replication cluster size".
+ * machines belonging to one party — one failure domain, one witness. A cross-party mesh has
+ * exactly one automatic path — a formation result carries the responder's strand addresses and
+ * `CadreNode.resolveCohortSeed` unions them in, seeding only the pair that just redeemed an
+ * invitation — and is otherwise hand-supplied addresses
+ * (`StrandConnectionOptions.bootstrapNodes`) or a hand-written dial (see the doc). Every party
+ * now holds its own membership identity, so the missing piece is placement, not identity:
+ * tracked as `backlog/feat-strand-party-diverse-placement`. Full reasoning:
+ * `docs/architecture.md` → "Replication cluster size".
  *
  * Raising or lowering it per strand is still the embedder's call
  * ({@link resolveStrandClusterSize}, `CadreNodeConfig.strandClusterSize`,
@@ -205,6 +247,17 @@ export declare const DEFAULT_STRAND_CLUSTER_SIZE = 4;
  * number, so the relaxed branch is reachable only by a cohort that is genuinely that small. See
  * Optimystic's `cluster/cluster-policy.ts` ("Why two size yardsticks, not one").
  *
+ * `assumedClusterSize` also carries the ADMISSION yardstick, and on the strand path it carries
+ * BOTH jobs today. Optimystic prefers a separately declared
+ * `repairCorroborationClusterSize` for the repair floor and falls back to this field only when
+ * nothing was declared — which, for a strand, is always: no per-strand serving count exists yet,
+ * so {@link strandClusterPolicy} declares nothing and the 2 here governs the repair floor
+ * exactly as the paragraph above describes, alongside the membership admission gate's
+ * low-confidence fallback. Do not raise it if a strand yardstick ever does get declared
+ * (`backlog/feat-strand-yardstick-from-serving-machines`): a strand shared between phones cannot
+ * promise that `ceil(0.75 x N)` of its machines are awake, and the admission gate would refuse
+ * its writes. Same reason and same warning as {@link CONTROL_CLUSTER_POLICY}'s.
+ *
  * The `satisfies` is load-bearing for the same reason as {@link CONTROL_CLUSTER_POLICY}'s.
  */
 export declare const STRAND_CLUSTER_POLICY: Readonly<{
@@ -229,3 +282,121 @@ export declare const STRAND_CLUSTER_POLICY: Readonly<{
  * this comment described. See `docs/architecture.md` → "Replication cluster size".
  */
 export declare function resolveStrandClusterSize(configured?: number): number;
+/**
+ * The repair yardstick to declare for a network, from the machines that SERVE that network and
+ * the breadth it replicates to.
+ *
+ * ## What this number does
+ *
+ * Optimystic's block repair (read-repair and reconcile) asks a block's cohort for the newest
+ * revision and trusts an answer only when enough independent peers agree on it. "Enough" is
+ * measured against a **declared** number rather than the peers currently visible — deliberately,
+ * because the visible set comes from unauthenticated routing, and a partition (or an attacker
+ * with routing influence) can shrink it. `clusterPolicy.repairCorroborationClusterSize` is that
+ * declaration; it moves the repair yardstick ALONE, leaving
+ * {@link CONTROL_CLUSTER_POLICY}'s / {@link STRAND_CLUSTER_POLICY}'s `assumedClusterSize` — the
+ * membership admission gate's yardstick — untouched. Optimystic resolves
+ * `repairCorroborationClusterSize -> assumedClusterSize -> clusterSize`, and a value that is not
+ * a positive integer falls THROUGH to the next term rather than being clamped, which is why the
+ * builders below refuse to pass a degenerate count at all.
+ *
+ * ## The arithmetic, and why the formula is what it is
+ *
+ * Three consumers read the resolved yardstick `N`. With `p` = cohort peers currently visible
+ * (self excluded) and Optimystic's `CORROBORATION_FLOOR` = 2:
+ *
+ * - **Repair corroboration floor** (`corroboratorCapacity` / `quorumSize` in
+ *   `db-p2p/src/cluster/quorum-restore.ts`): capacity is `max(p, N - 1)` and the floor is
+ *   `max(1, min(2, capacity))`. So `N = 1` and `N = 2` behave IDENTICALLY at every `p` — both
+ *   leave the floor at a single voter whenever `p <= 1`. `N >= 3` pins the floor at two
+ *   corroborators unconditionally, including when the view has shrunk to one peer. That is the
+ *   entire point of declaring. `N` above 3 raises nothing further (the `min(2, ...)` caps it), so
+ *   a larger declaration buys only honesty, never more repair.
+ * - **Commit freshness window** (`CoordinatorRepo.commitQuorumRulesOutRivals`): `fullCohortSize`
+ *   is `max(observed cohort, N)`, and a local commit arms the lazy read-repair window only when
+ *   `approvals > fullCohortSize / 2`. Here a larger `N` is NOT free — see the NOTE below.
+ * - The `repair-fault-tolerance` startup advisory — log text only.
+ *
+ * Hence `max(MIN_CLUSTER_SIZE, min(servingMachines, replicationBreadth))`, two clamps each for
+ * a stated reason:
+ *
+ * - **Capped at the replication breadth**, because a block only ever lives on
+ *   `min(breadth, machines serving)` machines. A party of eight running strands at
+ *   {@link DEFAULT_STRAND_CLUSTER_SIZE} has a cohort of four; declaring eight would make
+ *   `approvals > 4` unsatisfiable (a super-majority of four is three), so the freshness window
+ *   would never arm for any strand commit, forever — and buy nothing on repair, whose floor is
+ *   already at its cap at four. Declare the machines that can hold the block, not the machines
+ *   that exist.
+ * - **Floored at {@link MIN_CLUSTER_SIZE}**, so this can only ever RAISE the declared number
+ *   relative to the un-derived constants. A node that authorizes nobody is a genuine
+ *   founder-alone party, but it is equally a freshly seeded node whose membership rows have not
+ *   replicated yet, or one whose trusted-owner anchor is empty — and this node cannot tell those
+ *   apart. Declaring 1 is safety-neutral for repair (identical to 2 at every `p`, per the table
+ *   above) but it WOULD let a solo commit arm the freshness window, which is wrong if the "solo"
+ *   reading came from stale local records. Lifting that floor for a node that can prove it is
+ *   genuinely alone is separate work: `backlog/feat-solo-node-arms-its-own-freshness-window`.
+ *
+ * The yardstick is per-node and protects only that node's own reads — no node refuses another
+ * anything over it — so two members disagreeing (one has replicated a new `CadrePeer` row, one
+ * has not) is harmless by construction, not a race to close.
+ *
+ * ## The contract: callers pass machines that SERVE this network, or nothing at all
+ *
+ * Over-declaring is not merely wasteful, it is unsafe in the availability direction: at
+ * `N >= 3` the corroboration floor is pinned at two peers, so a cohort that can only ever
+ * field one peer can never repair at all (`cluster-fetch:no-quorum`), where a declaration of
+ * 2 would have let its single peer answer. So the number a caller passes must be the machines
+ * that actually serve the network in question — never a broader population that merely bounds
+ * it from above, and never a guess. A caller with no trustworthy count passes nothing, which
+ * the builders below turn into "declare no yardstick at all".
+ *
+ * The CONTROL network satisfies this from the party's enrolled machines, because every enrolled
+ * machine runs the control node by construction: there, enrolled IS serving
+ * ({@link controlClusterPolicy}, fed by cadre-core's `enrolled-machine-store.ts`). A STRAND does
+ * not — it launches only on machines whose embedder registered its sApp config
+ * (`CadreNode.addStrand`), so a closed strand shared by two machines of a three-machine party is
+ * served by two, and declaring three would be exactly the unsafe over-declaration above. No
+ * authenticated per-strand serving count exists yet, so strand nodes declare NOTHING and run the
+ * frozen {@link STRAND_CLUSTER_POLICY} — keeping the known, upstream-tracked single-voter
+ * exposure (`backlog/debt-read-repair-single-voter-corroboration`) rather than risking the
+ * cannot-repair-at-all failure. Building that count:
+ * `backlog/feat-strand-yardstick-from-serving-machines`.
+ *
+ * Both arguments must be positive integers; the builders below sanitize before calling, and
+ * a degenerate value here propagates (`NaN` in, `NaN` out) rather than being clamped.
+ *
+ * NOTE: on a party of three or more machines where fewer than half are awake, a commit reaches a
+ * downsized cohort and no longer arms the freshness window, so each such block's next read costs
+ * one cohort consult per read-repair window instead of trusting the local commit. That is the
+ * correct reading of the honest number — a commit that reached two of five machines genuinely
+ * does not rule out a rival quorum, and arming it today is a consequence of under-declaring — and
+ * the cost is bounded at one consult per block per window by the read path's solo-self-skip exit.
+ * NOT measured under load. If a large, mostly-hibernating party's read path ever shows up as
+ * slow, this is the first thing to look at.
+ */
+export declare function resolveRepairYardstick(servingMachines: number, replicationBreadth: number): number;
+/**
+ * {@link CONTROL_CLUSTER_POLICY} with the repair yardstick declared from `enrolledMachines`
+ * (see {@link resolveRepairYardstick}); the frozen base object ITSELF when the count is unknown,
+ * so the cold path is provably today's behaviour and identity assertions keep holding.
+ *
+ * The parameter keeps the name `enrolledMachines`, unlike {@link strandClusterPolicy}'s, because
+ * for the CONTROL network the two quantities are one and the same by construction: every
+ * enrolled machine runs the control node, so the machines enrolled in the party ARE the machines
+ * serving this network.
+ */
+export declare function controlClusterPolicy(enrolledMachines?: number): NonNullable<NodeOptions['clusterPolicy']>;
+/**
+ * {@link STRAND_CLUSTER_POLICY} with the repair yardstick declared from `servingMachines` — the
+ * machines that serve THIS strand — and this strand's own `clusterSize` (already resolved by
+ * {@link resolveStrandClusterSize}, so it is at least {@link MIN_CLUSTER_SIZE} and the formula's
+ * two clamps can never cross); the frozen base object ITSELF when the count is unknown.
+ *
+ * **The unknown path is the production path today.** No authenticated per-strand serving count
+ * exists, and the party's enrolled-machine count is emphatically not one (see
+ * {@link resolveRepairYardstick}'s contract section), so cadre-core passes nothing and every
+ * strand node runs the frozen constant. The parameter and its threading through
+ * `StartStrandConfig.servingMachines` stay in place as the seam that
+ * `backlog/feat-strand-yardstick-from-serving-machines` plugs a real count into.
+ */
+export declare function strandClusterPolicy(clusterSize: number, servingMachines?: number): NonNullable<NodeOptions['clusterPolicy']>;

package/dist/cluster-size.js

@@ -7,7 +7,12 @@ export const MIN_CLUSTER_SIZE = 2;
 /**
  * How many nodes each block of the **control** (cadre membership) database is replicated
  * to. Chosen to exceed any party's node count, so in practice every control block lands
- * on every member of the party.
+ * on every member of the party. Cohort width alone only delivers that for blocks written
+ * while the members were serving the network: a block committed while the writer was
+ * alone (the founder's genesis above all — collection headers are written once and never
+ * revised) has a cohort of one, and it reaches later joiners through cadre-core's
+ * peer-join block catch-up (`peer-join-backfill.ts`, membership-gated on the control
+ * network), not through this breadth.
  *
  * **Why full replication here.** Every control node reads the whole control database —
  * membership, peer addresses, the strand list — so a member left out of a block's cohort is
@@ -90,6 +95,20 @@ export const CONTROL_REPLICATION_BREADTH = 16;
  * is what the corroboration floor is asking. See Optimystic's `cluster/cluster-policy.ts`
  * ("Why two size yardsticks, not one") and `corroboratorCapacity` in `cluster/quorum-restore.ts`.
  *
+ * **`assumedClusterSize` is now the ADMISSION yardstick.** Since Optimystic grew
+ * `repairCorroborationClusterSize` (see {@link controlClusterPolicy}), the repair corroboration
+ * floor prefers that separate declaration and reads this field only when nothing was declared —
+ * the unknown-machine-count path, where the paragraph above still describes what it does. So on
+ * every path where the count IS known, the 2 here governs one consumer: the membership admission
+ * gate's low-confidence fallback path. Leaving
+ * it at 2 is deliberate and must NOT be "fixed" to match the repair yardstick. Raising it would
+ * make the gate demand `ceil(0.75 x N)` DECLARED peers before it admits a coordinator's peer
+ * set on that path, and a Cadre party is phones and laptops: it cannot promise that three
+ * quarters of its enrolled machines are awake at any moment, so a party of eight with three
+ * awake would stop transacting entirely. The repair yardstick has no such cost (it gates only
+ * whether this node believes a repair answer), which is exactly why the two are now separate
+ * numbers rather than one.
+ *
  * The `satisfies` is load-bearing: without it a mistyped or obsolete key would compile at both
  * this definition and every consumer, because TypeScript only excess-property-checks fresh
  * object literals, not a shared constant handed to `clusterPolicy`.
@@ -125,6 +144,13 @@ export const CONTROL_CLUSTER_POLICY = Object.freeze({
  * phones and laptops is the ordinary case, not the rare one. Six buys no more fault tolerance
  * than four and costs more overfetch (see the NOTE below).
  *
+ * **Exercised at four machines** by `integration-tests`'
+ * `scenarios/strand-two-party-two-machine.integration.ts`, the only scenario that runs a
+ * strand at this breadth. It shows a write committing while one of the four machines is off.
+ * It does NOT distinguish that from a commit on a cohort downsized to the three live holders
+ * (`allowDownsize` in {@link STRAND_CLUSTER_POLICY}), so the table row above remains reasoned
+ * arithmetic with an end-to-end existence proof under it, not a measurement of cohort width.
+ *
  * **It is also a correctness floor, not only a durability one — but the floor is measured in
  * machines, not in this number.** A node that has fallen behind and can ask exactly one peer
  * whether it is current has Optimystic accept that single answer as the cluster's truth
@@ -133,12 +159,24 @@ export const CONTROL_CLUSTER_POLICY = Object.freeze({
  * honestly reports the stale revision, the reader concludes it is current, and re-arms its
  * repair window forever. Measured on the control-DB replication scenario: 4 failures in 10 runs
  * at breadth 2, 0 in 20 at breadth 3, 0 in 10 at breadth 8. What lifts the floor off a single
- * voter is a *third machine*, not a wider target: `corroboratorCapacity` takes the cohort peers
- * actually present against {@link STRAND_CLUSTER_POLICY}'s declared `assumedClusterSize`, so a
- * three-machine strand needs two corroborators whatever this number says, and a two-machine one
- * cannot supply them however wide the target is. The underlying Optimystic behaviour is unfixed
- * — `backlog/debt-read-repair-single-voter-corroboration` — so a two-machine strand still takes
- * the exposure. Raising this number does not buy the way out; adding a machine does.
+ * voter is a third machine **or a declaration that there are three** — not a wider target.
+ * `corroboratorCapacity` takes the max of the cohort peers actually present and the declared
+ * repair yardstick, so raising *this* number (the replication breadth) does nothing: the
+ * yardstick is a separate declaration, and **a strand node makes no such declaration today**.
+ * The control network derives one from the party's enrolled machines
+ * ({@link controlClusterPolicy}), which is exact there because every enrolled machine runs the
+ * control node. A strand has no equivalent count: it launches only on machines whose embedder
+ * registered its sApp config, and reusing the party count over-declares — which is worse than
+ * declaring nothing, because at a yardstick of three or more the corroboration floor is pinned
+ * at two peers and a strand that can field only one could then never repair at all. So every
+ * strand runs {@link STRAND_CLUSTER_POLICY} unchanged and takes the single-voter exposure
+ * whenever its cohort view shrinks to one peer. The underlying Optimystic behaviour is unfixed
+ * (`backlog/debt-read-repair-single-voter-corroboration`) and the per-strand serving count that
+ * would let a strand declare honestly is `backlog/feat-strand-yardstick-from-serving-machines`
+ * — its seam ({@link strandClusterPolicy}'s `servingMachines`, threaded through
+ * `StartStrandConfig.servingMachines`) is already in place and fed by nothing. Raising this
+ * number does not buy the way out; a machine, or an honest declaration of the machines actually
+ * serving the strand, does.
  *
  * **Why not derived from the party or member count.** The strand's `Member` rows live *in* the
  * strand database, which runs on the strand libp2p node, whose cluster size is frozen at
@@ -156,11 +194,14 @@ export const CONTROL_CLUSTER_POLICY = Object.freeze({
  * cadre-driven path, the one that resolves a seed for itself, a strand's libp2p mesh is seeded
  * from exactly one party's own control-database peer rows (`CadreNode.resolveCohortSeed`), so
  * the cohort actually formed is that one party's machines. Four copies are, on that path, four
- * machines belonging to one party — one failure domain, one witness. A cross-party mesh is not
- * forbidden, only undiscoverable: it takes hand-supplied addresses
- * (`StrandConnectionOptions.bootstrapNodes`) or a hand-written dial (see the doc). Tracked:
- * `backlog/feat-strand-party-identity`. Full reasoning: `docs/architecture.md` →
- * "Replication cluster size".
+ * machines belonging to one party — one failure domain, one witness. A cross-party mesh has
+ * exactly one automatic path — a formation result carries the responder's strand addresses and
+ * `CadreNode.resolveCohortSeed` unions them in, seeding only the pair that just redeemed an
+ * invitation — and is otherwise hand-supplied addresses
+ * (`StrandConnectionOptions.bootstrapNodes`) or a hand-written dial (see the doc). Every party
+ * now holds its own membership identity, so the missing piece is placement, not identity:
+ * tracked as `backlog/feat-strand-party-diverse-placement`. Full reasoning:
+ * `docs/architecture.md` → "Replication cluster size".
  *
  * Raising or lowering it per strand is still the embedder's call
  * ({@link resolveStrandClusterSize}, `CadreNodeConfig.strandClusterSize`,
@@ -205,6 +246,17 @@ export const DEFAULT_STRAND_CLUSTER_SIZE = 4;
  * number, so the relaxed branch is reachable only by a cohort that is genuinely that small. See
  * Optimystic's `cluster/cluster-policy.ts` ("Why two size yardsticks, not one").
  *
+ * `assumedClusterSize` also carries the ADMISSION yardstick, and on the strand path it carries
+ * BOTH jobs today. Optimystic prefers a separately declared
+ * `repairCorroborationClusterSize` for the repair floor and falls back to this field only when
+ * nothing was declared — which, for a strand, is always: no per-strand serving count exists yet,
+ * so {@link strandClusterPolicy} declares nothing and the 2 here governs the repair floor
+ * exactly as the paragraph above describes, alongside the membership admission gate's
+ * low-confidence fallback. Do not raise it if a strand yardstick ever does get declared
+ * (`backlog/feat-strand-yardstick-from-serving-machines`): a strand shared between phones cannot
+ * promise that `ceil(0.75 x N)` of its machines are awake, and the admission gate would refuse
+ * its writes. Same reason and same warning as {@link CONTROL_CLUSTER_POLICY}'s.
+ *
  * The `satisfies` is load-bearing for the same reason as {@link CONTROL_CLUSTER_POLICY}'s.
  */
 export const STRAND_CLUSTER_POLICY = Object.freeze({
@@ -237,4 +289,155 @@ export function resolveStrandClusterSize(configured) {
     }
     return configured;
 }
+/**
+ * The repair yardstick to declare for a network, from the machines that SERVE that network and
+ * the breadth it replicates to.
+ *
+ * ## What this number does
+ *
+ * Optimystic's block repair (read-repair and reconcile) asks a block's cohort for the newest
+ * revision and trusts an answer only when enough independent peers agree on it. "Enough" is
+ * measured against a **declared** number rather than the peers currently visible — deliberately,
+ * because the visible set comes from unauthenticated routing, and a partition (or an attacker
+ * with routing influence) can shrink it. `clusterPolicy.repairCorroborationClusterSize` is that
+ * declaration; it moves the repair yardstick ALONE, leaving
+ * {@link CONTROL_CLUSTER_POLICY}'s / {@link STRAND_CLUSTER_POLICY}'s `assumedClusterSize` — the
+ * membership admission gate's yardstick — untouched. Optimystic resolves
+ * `repairCorroborationClusterSize -> assumedClusterSize -> clusterSize`, and a value that is not
+ * a positive integer falls THROUGH to the next term rather than being clamped, which is why the
+ * builders below refuse to pass a degenerate count at all.
+ *
+ * ## The arithmetic, and why the formula is what it is
+ *
+ * Three consumers read the resolved yardstick `N`. With `p` = cohort peers currently visible
+ * (self excluded) and Optimystic's `CORROBORATION_FLOOR` = 2:
+ *
+ * - **Repair corroboration floor** (`corroboratorCapacity` / `quorumSize` in
+ *   `db-p2p/src/cluster/quorum-restore.ts`): capacity is `max(p, N - 1)` and the floor is
+ *   `max(1, min(2, capacity))`. So `N = 1` and `N = 2` behave IDENTICALLY at every `p` — both
+ *   leave the floor at a single voter whenever `p <= 1`. `N >= 3` pins the floor at two
+ *   corroborators unconditionally, including when the view has shrunk to one peer. That is the
+ *   entire point of declaring. `N` above 3 raises nothing further (the `min(2, ...)` caps it), so
+ *   a larger declaration buys only honesty, never more repair.
+ * - **Commit freshness window** (`CoordinatorRepo.commitQuorumRulesOutRivals`): `fullCohortSize`
+ *   is `max(observed cohort, N)`, and a local commit arms the lazy read-repair window only when
+ *   `approvals > fullCohortSize / 2`. Here a larger `N` is NOT free — see the NOTE below.
+ * - The `repair-fault-tolerance` startup advisory — log text only.
+ *
+ * Hence `max(MIN_CLUSTER_SIZE, min(servingMachines, replicationBreadth))`, two clamps each for
+ * a stated reason:
+ *
+ * - **Capped at the replication breadth**, because a block only ever lives on
+ *   `min(breadth, machines serving)` machines. A party of eight running strands at
+ *   {@link DEFAULT_STRAND_CLUSTER_SIZE} has a cohort of four; declaring eight would make
+ *   `approvals > 4` unsatisfiable (a super-majority of four is three), so the freshness window
+ *   would never arm for any strand commit, forever — and buy nothing on repair, whose floor is
+ *   already at its cap at four. Declare the machines that can hold the block, not the machines
+ *   that exist.
+ * - **Floored at {@link MIN_CLUSTER_SIZE}**, so this can only ever RAISE the declared number
+ *   relative to the un-derived constants. A node that authorizes nobody is a genuine
+ *   founder-alone party, but it is equally a freshly seeded node whose membership rows have not
+ *   replicated yet, or one whose trusted-owner anchor is empty — and this node cannot tell those
+ *   apart. Declaring 1 is safety-neutral for repair (identical to 2 at every `p`, per the table
+ *   above) but it WOULD let a solo commit arm the freshness window, which is wrong if the "solo"
+ *   reading came from stale local records. Lifting that floor for a node that can prove it is
+ *   genuinely alone is separate work: `backlog/feat-solo-node-arms-its-own-freshness-window`.
+ *
+ * The yardstick is per-node and protects only that node's own reads — no node refuses another
+ * anything over it — so two members disagreeing (one has replicated a new `CadrePeer` row, one
+ * has not) is harmless by construction, not a race to close.
+ *
+ * ## The contract: callers pass machines that SERVE this network, or nothing at all
+ *
+ * Over-declaring is not merely wasteful, it is unsafe in the availability direction: at
+ * `N >= 3` the corroboration floor is pinned at two peers, so a cohort that can only ever
+ * field one peer can never repair at all (`cluster-fetch:no-quorum`), where a declaration of
+ * 2 would have let its single peer answer. So the number a caller passes must be the machines
+ * that actually serve the network in question — never a broader population that merely bounds
+ * it from above, and never a guess. A caller with no trustworthy count passes nothing, which
+ * the builders below turn into "declare no yardstick at all".
+ *
+ * The CONTROL network satisfies this from the party's enrolled machines, because every enrolled
+ * machine runs the control node by construction: there, enrolled IS serving
+ * ({@link controlClusterPolicy}, fed by cadre-core's `enrolled-machine-store.ts`). A STRAND does
+ * not — it launches only on machines whose embedder registered its sApp config
+ * (`CadreNode.addStrand`), so a closed strand shared by two machines of a three-machine party is
+ * served by two, and declaring three would be exactly the unsafe over-declaration above. No
+ * authenticated per-strand serving count exists yet, so strand nodes declare NOTHING and run the
+ * frozen {@link STRAND_CLUSTER_POLICY} — keeping the known, upstream-tracked single-voter
+ * exposure (`backlog/debt-read-repair-single-voter-corroboration`) rather than risking the
+ * cannot-repair-at-all failure. Building that count:
+ * `backlog/feat-strand-yardstick-from-serving-machines`.
+ *
+ * Both arguments must be positive integers; the builders below sanitize before calling, and
+ * a degenerate value here propagates (`NaN` in, `NaN` out) rather than being clamped.
+ *
+ * NOTE: on a party of three or more machines where fewer than half are awake, a commit reaches a
+ * downsized cohort and no longer arms the freshness window, so each such block's next read costs
+ * one cohort consult per read-repair window instead of trusting the local commit. That is the
+ * correct reading of the honest number — a commit that reached two of five machines genuinely
+ * does not rule out a rival quorum, and arming it today is a consequence of under-declaring — and
+ * the cost is bounded at one consult per block per window by the read path's solo-self-skip exit.
+ * NOT measured under load. If a large, mostly-hibernating party's read path ever shows up as
+ * slow, this is the first thing to look at.
+ */
+export function resolveRepairYardstick(servingMachines, replicationBreadth) {
+    return Math.max(MIN_CLUSTER_SIZE, Math.min(servingMachines, replicationBreadth));
+}
+/**
+ * A serving-machine count as a usable number, or `undefined` for "this node does not know".
+ * Anything that is not a positive integer is unknown rather than clamped: Optimystic itself
+ * treats a degenerate declaration as absent, so silently rounding one here would hide a caller
+ * bug behind a number nobody chose.
+ */
+function asKnownMachineCount(servingMachines) {
+    return servingMachines !== undefined
+        && Number.isInteger(servingMachines)
+        && servingMachines >= 1
+        ? servingMachines
+        : undefined;
+}
+/**
+ * {@link CONTROL_CLUSTER_POLICY} with the repair yardstick declared from `enrolledMachines`
+ * (see {@link resolveRepairYardstick}); the frozen base object ITSELF when the count is unknown,
+ * so the cold path is provably today's behaviour and identity assertions keep holding.
+ *
+ * The parameter keeps the name `enrolledMachines`, unlike {@link strandClusterPolicy}'s, because
+ * for the CONTROL network the two quantities are one and the same by construction: every
+ * enrolled machine runs the control node, so the machines enrolled in the party ARE the machines
+ * serving this network.
+ */
+export function controlClusterPolicy(enrolledMachines) {
+    const known = asKnownMachineCount(enrolledMachines);
+    if (known === undefined) {
+        return CONTROL_CLUSTER_POLICY;
+    }
+    return Object.freeze({
+        ...CONTROL_CLUSTER_POLICY,
+        repairCorroborationClusterSize: resolveRepairYardstick(known, CONTROL_REPLICATION_BREADTH)
+    });
+}
+/**
+ * {@link STRAND_CLUSTER_POLICY} with the repair yardstick declared from `servingMachines` — the
+ * machines that serve THIS strand — and this strand's own `clusterSize` (already resolved by
+ * {@link resolveStrandClusterSize}, so it is at least {@link MIN_CLUSTER_SIZE} and the formula's
+ * two clamps can never cross); the frozen base object ITSELF when the count is unknown.
+ *
+ * **The unknown path is the production path today.** No authenticated per-strand serving count
+ * exists, and the party's enrolled-machine count is emphatically not one (see
+ * {@link resolveRepairYardstick}'s contract section), so cadre-core passes nothing and every
+ * strand node runs the frozen constant. The parameter and its threading through
+ * `StartStrandConfig.servingMachines` stay in place as the seam that
+ * `backlog/feat-strand-yardstick-from-serving-machines` plugs a real count into.
+ */
+export function strandClusterPolicy(clusterSize, servingMachines) {
+    const known = asKnownMachineCount(servingMachines);
+    if (known === undefined) {
+        return STRAND_CLUSTER_POLICY;
+    }
+    return Object.freeze({
+        ...STRAND_CLUSTER_POLICY,
+        repairCorroborationClusterSize: resolveRepairYardstick(known, clusterSize)
+    });
+}
 //# sourceMappingURL=cluster-size.js.map

package/dist/cluster-size.js.map

@@ -1 +1 @@
-{"version":3,"file":"cluster-size.js","sourceRoot":"","sources":["../src/cluster-size.ts"],"names":[],"mappings":"AAEA;;;;GAIG;AACH,MAAM,CAAC,MAAM,gBAAgB,GAAG,CAAC,CAAC;AAElC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA6CG;AACH,MAAM,CAAC,MAAM,2BAA2B,GAAG,EAAE,CAAC;AAE9C;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA0CG;AACH,MAAM,CAAC,MAAM,sBAAsB,GAAG,MAAM,CAAC,MAAM,CAAC;IACnD,aAAa,EAAE,IAAI;IACnB,aAAa,EAAE,GAAG;IAClB,kBAAkB,EAAE,CAAC;CAC+B,CAAC,CAAC;AAEvD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAyEG;AACH,MAAM,CAAC,MAAM,2BAA2B,GAAG,CAAC,CAAC;AAE7C;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAgCG;AACH,MAAM,CAAC,MAAM,qBAAqB,GAAG,MAAM,CAAC,MAAM,CAAC;IAClD,aAAa,EAAE,IAAI;IACnB,aAAa,EAAE,GAAG;IAClB,kBAAkB,EAAE,gBAAgB;CACgB,CAAC,CAAC;AAEvD;;;;;;;;;;;;;;;GAeG;AACH,MAAM,UAAU,wBAAwB,CAAC,UAAmB;IAC3D,IAAI,UAAU,KAAK,SAAS,EAAE,CAAC;QAC9B,OAAO,2BAA2B,CAAC;IACpC,CAAC;IACD,IAAI,CAAC,MAAM,CAAC,SAAS,CAAC,UAAU,CAAC,IAAI,UAAU,GAAG,gBAAgB,EAAE,CAAC;QACpE,MAAM,IAAI,KAAK,CACd,qCAAqC,gBAAgB,6CAA6C,UAAU,EAAE,CAC9G,CAAC;IACH,CAAC;IACD,OAAO,UAAU,CAAC;AACnB,CAAC"}
+{"version":3,"file":"cluster-size.js","sourceRoot":"","sources":["../src/cluster-size.ts"],"names":[],"mappings":"AAEA;;;;GAIG;AACH,MAAM,CAAC,MAAM,gBAAgB,GAAG,CAAC,CAAC;AAElC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAkDG;AACH,MAAM,CAAC,MAAM,2BAA2B,GAAG,EAAE,CAAC;AAE9C;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAwDG;AACH,MAAM,CAAC,MAAM,sBAAsB,GAAG,MAAM,CAAC,MAAM,CAAC;IACnD,aAAa,EAAE,IAAI;IACnB,aAAa,EAAE,GAAG;IAClB,kBAAkB,EAAE,CAAC;CAC+B,CAAC,CAAC;AAEvD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA+FG;AACH,MAAM,CAAC,MAAM,2BAA2B,GAAG,CAAC,CAAC;AAE7C;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA2CG;AACH,MAAM,CAAC,MAAM,qBAAqB,GAAG,MAAM,CAAC,MAAM,CAAC;IAClD,aAAa,EAAE,IAAI;IACnB,aAAa,EAAE,GAAG;IAClB,kBAAkB,EAAE,gBAAgB;CACgB,CAAC,CAAC;AAEvD;;;;;;;;;;;;;;;GAeG;AACH,MAAM,UAAU,wBAAwB,CAAC,UAAmB;IAC3D,IAAI,UAAU,KAAK,SAAS,EAAE,CAAC;QAC9B,OAAO,2BAA2B,CAAC;IACpC,CAAC;IACD,IAAI,CAAC,MAAM,CAAC,SAAS,CAAC,UAAU,CAAC,IAAI,UAAU,GAAG,gBAAgB,EAAE,CAAC;QACpE,MAAM,IAAI,KAAK,CACd,qCAAqC,gBAAgB,6CAA6C,UAAU,EAAE,CAC9G,CAAC;IACH,CAAC;IACD,OAAO,UAAU,CAAC;AACnB,CAAC;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA2FG;AACH,MAAM,UAAU,sBAAsB,CAAC,eAAuB,EAAE,kBAA0B;IACzF,OAAO,IAAI,CAAC,GAAG,CAAC,gBAAgB,EAAE,IAAI,CAAC,GAAG,CAAC,eAAe,EAAE,kBAAkB,CAAC,CAAC,CAAC;AAClF,CAAC;AAED;;;;;GAKG;AACH,SAAS,mBAAmB,CAAC,eAAwB;IACpD,OAAO,eAAe,KAAK,SAAS;WAChC,MAAM,CAAC,SAAS,CAAC,eAAe,CAAC;WACjC,eAAe,IAAI,CAAC;QACvB,CAAC,CAAC,eAAe;QACjB,CAAC,CAAC,SAAS,CAAC;AACd,CAAC;AAED;;;;;;;;;GASG;AACH,MAAM,UAAU,oBAAoB,CAAC,gBAAyB;IAC7D,MAAM,KAAK,GAAG,mBAAmB,CAAC,gBAAgB,CAAC,CAAC;IACpD,IAAI,KAAK,KAAK,SAAS,EAAE,CAAC;QACzB,OAAO,sBAAsB,CAAC;IAC/B,CAAC;IACD,OAAO,MAAM,CAAC,MAAM,CAAC;QACpB,GAAG,sBAAsB;QACzB,8BAA8B,EAAE,sBAAsB,CAAC,KAAK,EAAE,2BAA2B,CAAC;KACtC,CAAC,CAAC;AACxD,CAAC;AAED;;;;;;;;;;;;GAYG;AACH,MAAM,UAAU,mBAAmB,CAClC,WAAmB,EACnB,eAAwB;IAExB,MAAM,KAAK,GAAG,mBAAmB,CAAC,eAAe,CAAC,CAAC;IACnD,IAAI,KAAK,KAAK,SAAS,EAAE,CAAC;QACzB,OAAO,qBAAqB,CAAC;IAC9B,CAAC;IACD,OAAO,MAAM,CAAC,MAAM,CAAC;QACpB,GAAG,qBAAqB;QACxB,8BAA8B,EAAE,sBAAsB,CAAC,KAAK,EAAE,WAAW,CAAC;KACtB,CAAC,CAAC;AACxD,CAAC"}

package/dist/index.d.ts

@@ -1,4 +1,4 @@
 export { connectToStrand } from './connect.js';
-export { MIN_CLUSTER_SIZE, CONTROL_REPLICATION_BREADTH, CONTROL_CLUSTER_POLICY, DEFAULT_STRAND_CLUSTER_SIZE, STRAND_CLUSTER_POLICY, resolveStrandClusterSize } from './cluster-size.js';
+export { MIN_CLUSTER_SIZE, CONTROL_REPLICATION_BREADTH, CONTROL_CLUSTER_POLICY, DEFAULT_STRAND_CLUSTER_SIZE, STRAND_CLUSTER_POLICY, resolveStrandClusterSize, resolveRepairYardstick, controlClusterPolicy, strandClusterPolicy } from './cluster-size.js';
 export { wrapStorageWithCache, disposeStorageCache } from './cached-storage.js';
 export type { StrandConnectionOptions, SereusPluginResult, StrandTransactor } from './types.js';

package/dist/index.js

@@ -1,4 +1,4 @@
 export { connectToStrand } from './connect.js';
-export { MIN_CLUSTER_SIZE, CONTROL_REPLICATION_BREADTH, CONTROL_CLUSTER_POLICY, DEFAULT_STRAND_CLUSTER_SIZE, STRAND_CLUSTER_POLICY, resolveStrandClusterSize } from './cluster-size.js';
+export { MIN_CLUSTER_SIZE, CONTROL_REPLICATION_BREADTH, CONTROL_CLUSTER_POLICY, DEFAULT_STRAND_CLUSTER_SIZE, STRAND_CLUSTER_POLICY, resolveStrandClusterSize, resolveRepairYardstick, controlClusterPolicy, strandClusterPolicy } from './cluster-size.js';
 export { wrapStorageWithCache, disposeStorageCache } from './cached-storage.js';
 //# sourceMappingURL=index.js.map

package/dist/index.js.map

@@ -1 +1 @@
-{"version":3,"file":"index.js","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,eAAe,EAAE,MAAM,cAAc,CAAC;AAC/C,OAAO,EACN,gBAAgB,EAChB,2BAA2B,EAC3B,sBAAsB,EACtB,2BAA2B,EAC3B,qBAAqB,EACrB,wBAAwB,EACxB,MAAM,mBAAmB,CAAC;AAC3B,OAAO,EAAE,oBAAoB,EAAE,mBAAmB,EAAE,MAAM,qBAAqB,CAAC"}
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,eAAe,EAAE,MAAM,cAAc,CAAC;AAC/C,OAAO,EACN,gBAAgB,EAChB,2BAA2B,EAC3B,sBAAsB,EACtB,2BAA2B,EAC3B,qBAAqB,EACrB,wBAAwB,EACxB,sBAAsB,EACtB,oBAAoB,EACpB,mBAAmB,EACnB,MAAM,mBAAmB,CAAC;AAC3B,OAAO,EAAE,oBAAoB,EAAE,mBAAmB,EAAE,MAAM,qBAAqB,CAAC"}