Define P2P session cancellation ownership

docs/adr/2026_08_replicator_capabilities_01_core_contract.md

@@ -331,6 +331,12 @@ initialisation, reset, lock, unlock, and resolution settle only after their
 defined remote write succeeds; a Rebuilder must not continue after an ignored
 mutation failure.
 
+`cancelled` means that the requested finite operation did not reach its normal
+completion boundary. It does not promise rollback. A provider may retain
+documents and checkpoints from batches which had already settled before the
+cancellation signal was observed, and a later operation resumes from that
+durable state.
+
 ### Distinguish observation from an identity result only when required
 
 Capability availability and operation results are separate. A supported
@@ -367,17 +373,22 @@ active -> retiring -> disposed -> replacement published
 ```
 
 Acquisitions wait for the transition and receive only the replacement. A fenced
-old handle cannot start work. Disposal stops Continuous activity, awaits work
+old handle cannot start work. Disposal stops Continuous activity, requests
-which cannot be cancelled, and settles or reports every operation owned by the
+cancellation of work which supports it, awaits work which cannot be cancelled,
-active adapter before publication. If bounded retirement cannot settle,
+and settles or reports every operation owned by the active adapter before
-replacement fails visibly and no new handle is published; the old handle is
+publication. If bounded retirement cannot settle, replacement fails visibly
-disposed when late work settles.
+and no new handle is published; the old handle is disposed when late work
+settles.
 
 The generic stop role is an idempotent request to stop a provider transfer after
 transport work begins. It does not promise cancellation of readiness checks,
 Security Seed acquisition, or external storage calls which do not consume a
-cancellation signal. P2P declares this role `not-implemented` until its lower
+cancellation signal. P2P implements this role through its room-session owner:
-level transfer and RPC work have a real stop path.
+the request aborts the current finite-operation scopes without closing the room
+or disabling later transfers. Its RPC request, incoming `reqSync`, and
+replication batch loop consume the same effective signal. An already-started
+atomic database operation may settle before cancellation completes, but no new
+batch is started afterwards.
 
 `getNewReplicator()` is not a general temporary-instance API. Setup and settings
 flows request narrow probes, such as connection, preferred-tweak, or isolated
@@ -426,7 +437,7 @@ support states.
 | Remote storage status                   | S       | S              | NA  |
 | Compromised-Chunk inspection            | S       | NA             | NA  |
 | Central-remote Security Seed            | S       | S              | NA  |
-| Request to stop active transfer         | S       | S              | NI  |
+| Request to stop active transfer         | S       | S              | S   |
 
 P2P unattended OneShot means a role exists which uses configured target names
 without opening a dialogue. Peer-room, watch, acceptance, and broadcast roles

docs/adr/2026_08_replicator_capabilities_02_p2p_service_lifecycle.md

@@ -65,7 +65,8 @@ A **P2P room session** is one active room membership and every resource whose
 validity depends on that membership:
 
 - the Trystero room, RPC actions, and `RpcRoom`;
-- its internal session epoch and cancellation signal;
+- its internal session epoch, session controller, and finite-operation
+  registry;
 - advertisement state and temporary peer decisions;
 - peer-bound RPC clients and remote database proxies;
 - connection, diagnostic, and platform-event subscriptions; and
@@ -147,11 +148,31 @@ dialogue. Persisted or automatic acceptance may authorise an unattended path;
 it cannot create local interaction authority implicitly.
 
 Every finite operation which needs a room obtains its own internal,
-session-epoch-bound demand. Acquiring another demand does not open another
+session-epoch-bound demand and operation controller. The operation consumes an
-session. Releasing it never closes a session still required by AutoStart,
+effective signal composed from the room session, its operation controller, and
-another finite operation, or another host consumer. Demand bookkeeping is
+any narrower caller or incoming-RPC signal. The room-session owner, rather than
-internal to the service and is not a general consumer contract; it cannot turn
+the adapter or UI consumer, owns every controller and operation settlement.
-a finite transfer into persistent transport policy.
+
+Acquiring another demand does not open another session. Releasing it never
+closes a session still required by AutoStart, another finite operation, or
+another host consumer. A request to stop active transfer aborts the registered
+finite-operation controllers but does not abort the room-session controller;
+the room remains usable and later transfers obtain fresh operation controllers.
+Retiring the room session aborts its session controller, which cancels every
+remaining child operation. Demand and controller bookkeeping is internal to the
+service and is not a general consumer contract; it cannot turn a finite transfer
+into persistent transport policy.
+
+Cancellation is cooperative and does not roll back durable work. Pull,
+requested push, and bidirectional synchronisation propagate the effective
+signal through the initiating RPC, incoming `reqSync`, the reverse database RPC
+calls, and the replication batch loop. An atomic PouchDB read or write which has
+already begun may settle. If a batch write has begun, the operation processes
+its successful writes and records the batch checkpoint only after every
+required revision has settled successfully. Cancellation before the write does
+not advance that checkpoint. The operation then reports a cancelled result and
+does not start another batch. Session retirement awaits that settlement before
+releasing RPC and room resources.
 
 ### Reconcile session settings atomically
 
@@ -173,19 +194,18 @@ client, or policy demand remains reachable after replacement.
 Reconciliation is serialised with room lifecycle operations:
 
 1. fence new session work;
-2. allow already-started finite transfers to settle, because the current lower
+2. abort the old session's finite-operation scopes and await their cooperative
-   level transfer may not yet have a real cancellation contract;
+   settlement;
 3. close and leave the old room in transport-owned order;
 4. open and validate the candidate session; and
 5. publish the replacement only after it has opened successfully.
 
-The service never exposes a partly initialised candidate or silently interrupts
+The service never exposes a partly initialised candidate or treats a cancellation
-a transfer without a cancellation contract. If bounded settlement or candidate
+request as rollback. If bounded settlement or candidate opening fails,
-opening fails, reconciliation reports the failure and publishes no mixed old
+reconciliation reports the failure and publishes no mixed old and new session.
-and new session. A candidate-open failure leaves one observable disconnected
+A candidate-open failure leaves one observable disconnected state with policy
-state with policy demands unsatisfied; it does not revive the fenced session or
+demands unsatisfied; it does not revive the fenced session or start an unbounded
-start an unbounded retry loop. A later lifecycle trigger or explicit connect may
+retry loop. A later lifecycle trigger or explicit connect may retry.
-retry.
 
 ### Order local database replacement across both owners
 
@@ -258,6 +278,15 @@ host must reject concurrent probing while the active service uses those
 sockets. It must not pause or close active relay sockets as a side effect of a
 short-lived check.
 
+The probe owns the same kind of lifetime controller as a room session, and
+`dispose()` aborts and settles its work before releasing its own room. This
+makes probe retirement bounded, but it does not turn process-global Trystero
+relay state into isolated state. Runtime acquisition therefore reuses or
+observes a compatible active session where the probe requires no second room,
+allows a separately owned logical room only when the effective relay binding is
+compatible, and otherwise returns a blocked result. It does not silently retire
+the active service to make an incompatible probe possible.
+
 ### Keep provider composition explicit
 
 The stable P2P service can be composed even when P2P is not the selected main

docs/adr/2026_08_replicator_capabilities_03_migration_plan.md

@@ -116,10 +116,14 @@ Preserve RTC diagnostics through `P2PDiagnostics`, rather than retaining
 `rawHost`. The compatibility facade may delegate during this stage, but new
 consumers cannot receive it.
 
-Separate active-adapter release, room-session leave, and the stop request. Keep
+Separate active-adapter release, room-session leave, and the stop request.
-the P2P stop role `not-implemented` until a lower-level operation exists. Add
+Implement the lower-level cooperative cancellation path before declaring the
-internal session-demand ownership for finite operations and policy-held
+P2P stop role supported: caller abort through RPC request cancellation,
-AutoStart without exposing that bookkeeping as a general consumer API.
+incoming-handler signal propagation, signal-bound reverse database RPC calls,
+and safe batch-boundary termination in `replicateShim`. Add room-session and
+operation controllers beside internal session-demand ownership for finite
+operations and policy-held AutoStart, without exposing that bookkeeping as a
+general consumer API.
 
 Add ownership regressions immediately before implementation:
 
@@ -132,7 +136,19 @@ Add ownership regressions immediately before implementation:
   advertisements do not;
 - local database replacement retires database-bound feeds and publication;
 - repeated replacement does not retain platform-event subscriptions;
-- policy-only change waits for an already-started finite transfer to settle;
+- an active-transfer stop aborts finite operations without closing the room,
+  while a later operation can use the same room;
+- room retirement aborts both locally initiated and incoming `reqSync` work,
+  waits for an already-started atomic database operation to settle, and starts
+  no later batch;
+- RPC cancellation, timeout, peer departure, and room close abort a
+  cancellation-aware handler rather than only discarding its eventual result;
+- the inbound request context exists before request admission begins, so a
+  cancellation received while admission waits cannot be lost;
+- cancellation retains already-settled documents and checkpoints and reports
+  `cancelled`, rather than claiming rollback or completion;
+- a per-document batch-write failure does not advance the replication
+  checkpoint past the failed revision;
 - explicit disconnect suppresses AutoStart and relay reconnection until
   explicit connect; and
 - database replacement fences both active-provider and P2P work before