Spindel Engine Formalism

A formal account of the spindel reactive engine: its algebraic properties, flow-chart diagrams, and a correctness argument per reactive composition.

This doc does not re-explain the model — concepts.md teaches signals, spins, track/await, checkpoints, and the drain queue, and engine.md covers the implementation (state shape, addressing, CPS/trampoline, executors, overlay backend). This doc formalizes what those teach: every claim is stated as a law and classified Holds (with the enforcing code), Conjectured (plausible but not enforced/tested), or scoped/fragile (with the boundary noted). Citations are file:line against the source tree.

---

Part 1 — Algebraic Properties

1.1 track as a comonad

Claim. track exposes a comonadic structure: the value it delivers is an Interval (old, new, deltas) — a context-carrying value, not a bare value. The comonad is "a value in the context of its own history."

Mapping to comonad operations:

• extract : W a → a is @interval / (:new interval) — discard the context, take the current value (interval.cljc:97-98).
• duplicate : W a → W (W a) corresponds to re-tracking: the body re-runs and track again delivers an interval, whose old is the previous new. The persistence of the track continuation (:kind :track, never reaped — clear-all-await-continuations! only reaps :await-once/:external-await, engine/impl/simple.cljc:1597-1600) is what makes the comonadic co-bind well-defined: the body is a co-Kleisli arrow W a → b that the engine re-invokes on each new context.

Counit / identity law — extract ∘ duplicate = id. The first track call returns (iv/->Interval old snapshot deltas) synchronously (effects/track.cljc:197); @ of it yields snapshot, the live value. Holds structurally.

Comonad coherence across re-runs. On a signal change the engine resumes the earliest track cont and truncates later ones (truncate-stale-conts!, engine/impl/simple.cljc:329-361); the body slice re-runs from that track point with :slice-state restored. The "context" delivered on resume comes from :on-resume → get-track-value-if-newer (effects/track.cljc:71-103), which honours generation so a stale resume delivers deltas = nil rather than replaying consumed deltas. Holds, modulo the generation caveat in 1.7.

Status: Holds (structural). The comonad laws are not mechanized; the structure is real and the engine's persistence + interval delivery realize it. Conjectured: a full categorical proof would need track to be referentially transparent in its argument, which it is (it dispatches solely on SignalRef identity, effects/track.cljc:209-212).

1.2 await as a monad

Claim. await is monadic sequencing over Result. The CPS transform makes a spin body a sequence of await-bind steps; each await is >>= : M a → (a → M b) → M b where the continuation is the a → M b arrow.

• Return — a spin that resolves synchronously is return v: the body invokes resolve directly (spin/core.cljc:311-342).
• Bind — await child suspends the parent; when child completes, :resume-fn pattern-matches the child's Result and routes :ok payload to the cont's :resolve-fn, :error payload to :reject-fn (effects/await.cljc:129-132). This is the Either/Result monad's bind: errors short-circuit.

Left identity — return a >>= f = f a. An await of an already-completed child takes the fast path and resumes inline with the cached value (effects/await.cljc:165-190). Holds.

Right identity — m >>= return = m. A spin body whose last form is (await child) resolves with the child's value unchanged (the outer resolve is called with the child payload). Holds.

Associativity — (m >>= f) >>= g = m >>= (λx. f x >>= g). CPS nesting is associative by construction: ioc/invert (spin/cps.cljc:136) produces a right-nested continuation chain regardless of how awaits are grouped in source. Holds (structural).

Error short-circuit — the Result monad's defining property. A child :error routes to :reject-fn → abort-spin-chain! (spin/core.cljc:345-381, :524-546), which caches the error and marks observers dirty so they rediscover it. Holds.

Rendering consequence. Because errors short-circuit every ancestor bind, a single body exception rejects the whole chain up to the root render effect, which produces no vdom — the UI freezes at the last resolved frame with one :render/error log line. Semantically correct, observationally near-silent (the simmis AsOfDB episode: a contains? probe threw inside the page-editor body and the app appeared to ignore all updates). Dev builds now surface this visibly (dom/render.cljc::show-render-error-overlay!).

Status: Holds (structural). Not mechanized; CPS construction guarantees the monad laws the way any CPS encoding of a monad does.

1.3 The typed delta algebra laws

Each algebra is a monoid (D, ·, id) with -apply-deltas, -compose-deltas, -empty-deltas (incremental/algebra.cljc:42-63), subject to three laws (stated algebra.cljc:13-17).

Identity — apply(t, id) = t.

• ScalarAlgebra — id = ::no-change; -apply-deltas returns old unchanged for it (algebra.cljc:186-188). Holds.
• SequenceAlgebra — id = (empty-diff size) (all-zero/empty fields, sequence_algebra.cljc:70-74); apply-pipeline of an empty diff is a grow-0/shrink-0/identity-perm/empty-change no-op (sequence_algebra.cljc:91-114). Holds.
• MapAlgebra — id = {:assoc {} :dissoc #{} :update {}}; apply-pipeline over empty fields is identity. Holds.

Application homomorphism — apply(t, d₁·d₂) = apply(apply(t,d₁), d₂).

• Scalar — last-write-wins compose; apply(apply(t,d1),d2) and apply(t,d2) agree because compose returns d2 unless d2 is no-change (algebra.cljc:193-198). Holds.
• Sequence — compose-pair (sequence_algebra.cljc:136-172) is carefully built so the composed diff reproduces the two-step effect; the namespace docstring states this and a property test is referenced (algebra.cljc and sequence_algebra.cljc:38-48). Holds (conjectured complete; relies on the property test).
• Map — compose recurses one level into inner algebras (map_algebra.cljc:37-44). Holds for one nesting level. Scope boundary: sequence-nested lift/lower is deferred (algebra.cljc:259-264) — the :change field has no (:nested algebra delta) variant yet, so the homomorphism is not claimed for sequence-nested paths.

Associativity of compose — (d₁·d₂)·d₃ = d₁·(d₂·d₃).

• Scalar — trivially associative (last-write-wins). Holds.
• Sequence — permutations form a group, the other fields are additive monoids; sequence_algebra.cljc:36-48 claims the quotient monoid is associative. Holds (conjectured, property-tested).
• Map — per-key folding; associative because each field's merge is associative. Holds (conjectured).
• compose-deltas variadic left-folds (algebra.cljc:118-125); the fold is well-defined iff the binary op is associative.

nil vs [] contract. nil = uncertainty, [] = verified no-change (interval.cljc:30-49). no-change? checks both old = new and the delta being the algebra's identity (interval.cljc:186-219). Holds where producers obey it — fragile because it is a convention, not a type: nothing prevents a combinator from emitting nil when it could emit []. The docstring explicitly warns "Combinators MUST NOT conflate them via (seq deltas) alone."

Size-composability invariant (Sequence). -compose-deltas throws if size-after(d1) ≠ size-before(d2) (sequence_algebra.cljc:395-401). This makes the sequence algebra a partial monoid — composition is defined only for size-compatible pairs. Holds (enforced by throw); worth noting it is not a total monoid.

1.4 Glitch-freedom

Law. Within one drain, a spin re-runs at most once per generation, and observers run after their dependencies (topological order). This law is scoped to the synchronous sub-graph — see the async caveat.

• Topological dispatch. :signal-change resolves observers via rtp/ordered-observers → topological-sort (Kahn's algorithm, engine/impl/graph.cljc:30-81) over the transitive observer closure. Holds for the synchronous portion of the graph.
• Descendant filtering. compute-descendant-observers (engine/impl/simple.cljc) finds observers created by another observer; delegated-descendants keeps those the creator's resume really re-creates — the creator's resume cont must PRECEDE the child's registration — its :order is within the creator's continuation count at the registration (:creator-cont-mark, an O(1) watermark recorded by register-spin!). :order is count-based, so this relies on the continuations counted in the mark outliving the registration; where a cont is removed in between, the test errs either way — an extra independent resume, or a wrongly delegated child left :dirty (re-run when next reached) instead of eagerly re-run; no violation has been constructed — or the creator must itself be a delegated descendant (it re-runs from scratch). Those are filtered and invalidated in the same step (invalidate-node), so deferring the re-run to the parent never leaves a stale :clean cache behind (3.4). A child created before the creator's resume point — (let [c (spin (track a) …) x (await c) v (track a)] …) — is not re-created by that resume, so it is resumed like an independent observer (it used to be filtered too, and the parent glitched: [1 2] for [2 2]). Holds.
• Escalation. find-root-await-ancestor escalates a child observer to its root await-ancestor so child-fires-then-parent-refires double-execution is avoided (engine/impl/simple.cljc:463-490, :561-588). Holds.
• Per-generation dedup. :spin-completion keys [parent child generation] in (:resumed-conts batch) and skips an already-resumed triple (engine/impl/simple.cljc:653-679). Track deltas are deduped per-observer by :consumed-generation (1.7).
• Async caveat. The docstring at engine/impl/simple.cljc:511-526 is explicit: glitch-freedom holds only within the synchronous portion. A body suspended on a :deferred-delivery / :mailbox-post resolves in a later drain pass — async branches are only eventually consistent. A diamond where one arm is async and one is sync can transiently show a glitch (the sync arm updates, the async arm lags). This is a deliberate design boundary, not a bug, but it means the law is scoped, not absolute.

Status: Holds for the synchronous sub-graph; explicitly scoped (not a global law) across async boundaries.

1.5 Determinism / replay

Law. Re-executing the same code path produces the same spin ids, hence cache hits land on the same nodes — enabling fork/restore.

• Computation spins — ids minted by next-address! from (source-loc, chain-head) (addressing.cljc:133-155). The chain-head is per-spin, seeded at body start with body-start-chain-head (addressing.cljc:74-87) and restored at resume from :slice-state :chain-head. The same call sequence in the same body produces the same id sequence. Holds.
• The computation-vs-resource split is the replay-membership rule. Only :computation spins are in the replay set: their ids are deterministic and register-spin! can serve cached results under rebuild. :resource spins get gensym ids (spin/core.cljc:428) and are not replayable — register-spin! always re-runs them (engine/impl/simple.cljc:1493). This is correct: a sleep/mailbox body is effectful and one-shot; replaying it would re-arm the timer. Holds.
• Rebuild mode — Spin invoke Case 1a (spin/core.cljc:231-253) executes the body for side effects (re-creating nested spins, re-registering conts) while returning the cached value. Relies on the per-spin chain-head being re-seeded (spin/core.cljc:240). Holds.
• Usage contract — determinism depends on the body being a pure function of its captures and source location. A body that mints spins inside a data-dependent loop without with-key (addressing.cljc:194-216) will produce colliding ids across iterations. The engine does not detect this; it is a usage contract.

Status: Holds; the kind split makes the replay set precisely the content-addressed spins.

1.6 The B-gate (idempotence of re-registration)

Law. Re-registering a :computation spin re-runs its body iff its captured environment changed; otherwise re-registration is a no-op on cache validity (idempotent).

register-spin! (engine/impl/simple.cljc:1476-1501): on re-registration of an existing node, it re-runs (marks {:completed? false} + dirty) when

(or (= kind :resource)
    (captures-changed? old-captures new-captures))

captures-changed? (engine/impl/simple.cljc:1441-1452) returns true if the key set differs or any value is not identical? to its prior.

• Idempotence — re-registering with identical? captures leaves the node :clean and :completed? true: await/@spin serve the cache. Holds.
• Re-registration with changed captures invalidates the node's run generation (invalidate-node): any run of the old closure — in flight, or re-completed later by a track resume — closes over the old captures, and its result must not be certified :clean (3.4, generation rule). Only a fresh body entry certifies again. Holds.
• identical? not = — O(1), never throws; Clojure structural sharing keeps unchanged persistent values identical?. The cost is an occasional needless re-run when a value is rebuilt =-equal but not identical?. This makes the gate a sound over-approximation: it may re-run when it didn't need to, but it never skips a re-run that was needed. Holds (conservative).
• :resource always re-runs — correct, since resource bodies are effectful (1.5).
• First registration — new node, records captures, no re-run gate applies. Holds.

This is the formal content of the stale_lexical_scope_test regression (test/.../stale_lexical_scope_test.cljc): a nested (spin …) form re-evaluated by a re-running parent yields a fresh closure; the B-gate re-runs it iff its captured environment actually changed.

Status: Holds; sound conservative idempotence.

1.7 Per-observer delta-once delivery

Law. Each observer of a signal sees each generation's deltas exactly once.

A track cont captures the signal's :generation at registration as :consumed-generation (effects/track.cljc:128-129, :182-186). On resume, get-track-value-if-newer returns deltas only if current-generation > consumed-generation, else delivers (Interval new new nil) (effects/track.cljc:71-103).

Open edge case. The :consumed-generation is captured once at cont creation and is not advanced when the cont fires. get-track-value-if-newer compares against the original captured generation. On resume the body slice re-runs and track is called again, creating a new cont with the then-current generation — so in the normal track-resume path the consumed generation does advance via cont replacement. But the persistent track cont kept by truncate-stale-conts! for :order-less conts is re-tracked via track-signal-dep!, and that re-track does not refresh :consumed-generation. Conjectured latent issue: a spin tracking signals A and B, resumed on B, keeps A's cont with A's old consumed-generation; if the body slice that re-tracks A is the truncate re-track (not a body re-run), the generation is stale. The multi_track_observer_test pins the observer-set survival but does not assert delta-once for the kept cont.

Status: Holds for the common path; conjectured edge case around truncate-re-tracked conts — a targeted test is recommended.

1.8 Commutativity of independent signal changes

Law. Two signal changes to independent sub-graphs produce the same final state regardless of order.

Each :signal-change is processed as its own event with its own batch (engine/impl/simple.cljc:532-535); FIFO ordering of the queue means two enqueued signal changes drain sequentially. For disjoint observer sets the two drains touch disjoint nodes, so the result is order-independent. Holds (conjectured) — not mechanized, but follows from disjointness + the per-event batch isolation.

For overlapping sub-graphs order matters and the engine makes no commutativity claim — last-write-wins on the shared spin's cache. The batch macro (signal.cljc:21-55) coalesces multiple swaps into one :signal-change per distinct signal in insertion order, which is a determinism aid, not a commutativity guarantee.

1.9 Resume idempotence

Law. resume-body! applied to the same (spin, cont) is safe to attempt more than once without double-executing side effects.

resume-body! (engine/impl/simple.cljc:411-…) itself is not idempotent — it unconditionally re-runs the body slice. Idempotence is provided by the layers around it:

• :spin-completion dedup — [parent child generation] triple guard prevents a second resume of the same await cont in one batch (engine/impl/simple.cljc:653-679).
• Cancellation gate — for external-await, truncate-stale-conts! fires :cancel! before dissociating a stale cont (engine/impl/simple.cljc:355-356); the orphaned external closure then no-ops via cancellable-external-pair's gate (effects/await.cljc:322-463). This is precisely the cont_cancellation_test regression: a deferred's :pending holds both the orphaned and the new resolve closure; only the new one advances. Holds.
• try-claim-execution! — atomic cache+running check prevents two :spin-execution events from both running a cold body (engine/impl/simple.cljc:1852-1899). Holds.

Status: Holds — idempotence is enforced by dedup + cancellation gate + claim, not by resume-body! itself. resume-body! is the one resume path, taking a single mode ∈ {:track :await} argument. The mode is a case with no :else, so an illegal mode throws rather than silently mis-resuming; the per-mode data (:signal-id to re-track for :track, :resume-fn for :await) is carried by the continuation itself.

---

Part 2 — Flow-Chart Formalism

2.1 The drain cycle

flowchart TD
    A[external change: swap! / completion / delivery] --> B[enqueue-event! → :engine/pending]
    B --> C[trigger-drain! dispatches drain-events! on the executor]
    C --> D{CAS :engine/draining? false→true}
    D -- lost --> E[another drainer owns lock; return 0]
    D -- won --> F[loop: dequeue-event!]
    F --> G{event?}
    G -- yes --> H[process-event! - may enqueue cascade events]
    H --> F
    G -- no --> I[release :engine/draining? lock]
    I --> J{pending non-empty AND context alive?}
    J -- yes --> C
    J -- no --> K[drain complete]

2.2 Signal-change propagation

flowchart TD
    A["swap! sig f"] --> B["swap-signal*-explicit:
       snapshot←clean-value, old-snapshot←prev,
       deltas←get-deltas, generation←inc"]
    B --> C["enqueue {:type :signal-change :id sid}"]
    C --> D["process-event! :signal-change"]
    D --> E["create-batch (generation++)"]
    E --> F["clear-all-await-continuations! (reap :await-once/:external-await)"]
    F --> G["ordered-observers: topological sort of transitive closure"]
    G --> H["keep observers with a live track cont for sid"]
    H --> I["filter descendant observers (invalidate-node them); escalate await-children to root"]
    I --> J["for each independent observer: resume-body! mode :track"]
    J --> K["truncate-stale-conts! (order > track cont)"]
    K --> L["restore :slice-state (bindings, chain-head, tracking)"]
    L --> M["body slice re-runs from track point"]
    M --> N{body resolves synchronously?}
    N -- yes --> O["outer resolve → cache-result! → enqueue :spin-completion"]
    N -- no --> P["body suspends on await; control returns"]
    O --> Q["cache-result! marks observers dirty → cascade"]
    D --> R["clear :engine/current-batch at handler exit"]

2.3 The await-cascade

flowchart TD
    A["child spin completes: outer resolve fires"] --> B["spin-cache-result! child (ok value)"]
    B --> C["graph-commit-deps! child"]
    C --> D["enqueue-completion-event! {:type :spin-completion :id child}"]
    D --> E["process-event! :spin-completion"]
    E --> F["parents = subscriptions[[:spin/complete child]]"]
    F --> G["for each parent cont (not already-resumed in this batch)"]
    G --> H["resume-body! mode :await
             (no truncate, no dirty-remark, no created-spins clear;
              uses the cont's own :resume-fn)"]
    H --> I[":resume-fn pattern-matches child Result:
            :ok→resolve-fn, :error→reject-fn"]
    I --> J["parent body slice continues past (await child)"]
    E --> K["propagate-await-dirty!: mark COMPLETED non-suspended parents dirty"]

2.4 Track-resume (comonad) vs await-resume (monad)

flowchart LR
    subgraph TRACK["resume-body! mode :track — COMONADIC"]
      T1["signal change"] --> T2["resume EARLIEST track cont"]
      T2 --> T3["truncate stale conts + re-track skipped deps;
                 re-track cont's :signal-id;
                 mark {:completed? false :running? true} + dirty;
                 clear :created-spins"]
      T3 --> T4["body slice RE-RUNS from track point;
                 fresh Interval (old,new,deltas) delivered"]
      T4 --> T5["cont is PERSISTENT — survives, will fire again"]
    end
    subgraph AWAIT["resume-body! mode :await — MONADIC"]
      A1["child completes"] --> A2["resume the await cont"]
      A2 --> A3["NO truncation, NO signal re-track,
                 NO dirty/running re-mark,
                 NO :created-spins clear"]
      A3 --> A4["still-suspended body slice ADVANCES past await;
                 child Result bound in via the cont's :resume-fn"]
      A4 --> A5[":await-once reaped at generation bound;
                 :await-reactive persists"]
    end

The two columns are exactly the comonad/monad distinction, and the mode argument names which one. A :track resume re-runs (comonadic co-bind: re-invoke the co-Kleisli arrow on a new context); an :await resume advances (monadic bind: feed the bound value into the rest). resume-body! derives all the per-mode behavior from mode — the four behaviors are not a 4-tuple of caller-supplied flags.

2.5 Computation-spin lifecycle

stateDiagram-v2
    [*] --> Registered: make-spin (3-arity) → register-spin! :computation
    Registered --> Running: (spin r e) invoke Case 2 / :spin-execution / await slow path
    Running --> Completed: outer resolve → cache-result! (:clean, completed?, running?=false)
    Running --> StaleRun: invalidated in flight → invalidate-node (:dirty, :invalid-gen = :run-gen)
    StaleRun --> Dirty: outer resolve OR reject → cache-result! (result kept, NOT certified :clean)
    StaleRun --> Running: fresh body entry (enter-body! → new :run-gen; prologue cancels the stale run, hands its callbacks over)
    Running --> Suspended: body returns the incomplete sentinel (await/track)
    Suspended --> Running: resume-body! (cont fires)
    Completed --> Dirty: dependency changed → mark-dirty! / cache-result! cascade
    Completed --> Completed: re-register with identical? captures (B-gate no-op)
    Dirty --> Running: re-run (track resume / :spin-execution / B-gate re-register / await slow path)
    Completed --> [*]: GC + no observers + no live signal cont → full-cleanup-spin!

Body-entry prologue invariant. Every Registered/Dirty → Running transition that invokes the cps-fn from scratch MUST run the shared prologue simple/enter-body! (mark-running!, clear-created-spins!, clear-prior-body-conts!, seed-body-chain-head!; rebuild paths use the :rebuild? variant that leaves run-state untouched). The prologue maintains the invariant the dispatch rules silently assume: at most one body-generation of continuations exists per spin. Earliest-cont selection (§2.4 T2) and truncate-stale-conts! (§3.4) are sound only under it — a surviving earlier-generation cont always wins dispatch, and resuming it truncates the live body's conts and re-runs a frozen lexical closure (the zombie-repaint bug). The enumeration of body-entry sites is normative: Spin -invoke Case 2, await-spin's slow path (the one entry that calls .-spin-fn directly — it bypassed the prologue until 71d5801), the -invoke rebuild branch, and deref-spin's JVM rebuild path; :spin-execution events route through -invoke. Any NEW direct body invocation must call enter-body! — see its docstring's call-site list.

2.6 Resource-spin lifecycle

stateDiagram-v2
    [*] --> Armed: make-spin (1/2-arity) → register-spin! :resource (gensym id)
    Armed --> Fired: body invoked once (sleep timer / parallel children / deferred)
    Fired --> Resolved: resolve/reject called → cache-result!
    Resolved --> Disposed: GC cleanup (no replay path, always re-run on re-register)
    Armed --> Armed: re-register → ALWAYS re-run (kind=:resource)

2.7 Fork / restore

flowchart TD
    A[parent ExecutionContext - AtomBackend] --> B[fork-context]
    B --> C["OverlayBackend over parent backend"]
    C --> D["fork-local state: :track-subscriptions + :await-conts
             COPIED, :engine/pending [], fresh draining flag"]
    C --> E["shared-path reads fall through to parent
             (:nodes, :engine/cancelled-tokens depth-1 CoW)"]
    D --> F[fork drains independently; writes land in overlay]
    A --> G[snapshot-context: freeze - :running=nil]
    G --> H[restore-snapshot: revive, drive drain-events! directly]

---

Part 3 — Correctness per Reactive Composition

For each composition: the invariant that must hold, and the argument the engine preserves it (or where it is fragile).

3.1 Nested spins (spin-in-spin)

Invariant. A (spin …) form lexically inside another spin's body, re-evaluated when the parent re-runs, must observe its current captured lexical scope — no stale closure.

Argument. The parent re-run re-evaluates the inner (spin …) form, producing a fresh cps-fn closure over the new locals and a fresh :captured-locals map. register-spin!'s B-gate (engine/impl/simple.cljc:1489-1497) compares :captured-locals with identical? and re-runs the inner spin iff a capture changed. The inner spin's id is stable across parent re-runs (per-spin chain-head re-seeded at parent body start, addressing.cljc:121-127), so the re-evaluation lands on the same SpinNode — re-registration, not a new node. Holds — this is the stale_lexical_scope_test / nested_spin_invalidation_test regression, fixed by the B-gate.

3.2 track inside a re-running spin

Invariant. Each track call re-registers a persistent cont; the spin remains an observer of every signal it tracks, across re-runs.

Argument. track-signal-dep! eagerly registers the observer at track time (engine/impl/simple.cljc:1335-1384). On a body re-run the new track call re-registers; record-deps! at body completion prunes only signals tracked last run but not this run (engine/impl/simple.cljc:1063-1084). Fragile point — see 1.7 and 3.4: the :slice-state :tracking snapshot must include the just-tracked signal, else record-deps! would commit deps missing it (effects/track.cljc:146-159 documents this fix; multi_track_observer_test is the regression). Holds post-fix.

3.3 await inside a spin (await-cascade)

Invariant. When an awaited child completes, the parent resumes exactly once per completion, with the child's Result correctly routed (:ok→value, :error→propagated).

Argument. :spin-completion dispatch resumes the parent's await cont via resume-body! in :await mode, the cont's :resume-fn doing the monadic match (effects/await.cljc:129-132). The [parent child generation] dedup prevents a double resume in one batch. :await mode does no truncation, no dirty/running re-mark, and no :created-spins clear, because it advances a still-suspended body rather than re-running it — wiping :created-spins would drop the awaited child itself. Holds.

3.4 track + await in one body (continuation :order truncation)

Invariant. A body (track A) … (await X) …: if A changes while suspended on X, the body re-runs from the track A point and the stale await X cont must be cancelled so X's eventual completion does not also advance the orphaned slice.

Argument. truncate-stale-conts! drops every cont with :order > the resumed track cont, firing :cancel! first (engine/impl/simple.cljc:329-361). For external resources the cancellation gate (cancellable-external-pair) makes the orphaned closure a no-op (effects/await.cljc:322-463). For a child Spin await, the truncated cont simply ceases to exist, so the child's :spin-completion finds no subscriber for the old cont. The re-running body re-issues a fresh await X cont. Holds — this is the cont_cancellation_test suite (deferred, mailbox, fork-isolation).

Generation rule (no stale certification). The truncation above covers the case where the spin is resumed at its track point. A spin can also be invalidated without being resumed, while a run of it is in flight:

• a :signal-change filters it out as a descendant observer — its re-run is delegated to the ancestor that re-creates it (1.4);
• a re-running parent re-registers it with changed captures (B-gate, 1.6).

Neither truncates the in-flight run, and they must not: an awaiter outside the re-running subtree may be parked on it, and would be stranded. The invariant is instead: a run may cache its result as :clean only if nothing invalidated the node since the run began. It is enforced with run generations (engine/impl/simple.cljc):

• every fresh body entry (enter-body! → mark-running!) stamps the node with a new, process-wide unique :run-gen and returns it; the run's resolve/reject continuations (-invoke Case 2, the await slow path) hand it back to cache-result!;
• invalidate-node records the generation it invalidates (:invalid-gen) and marks the node :dirty; :completed? is left alone (it says the body finished — the GC reap predicate needs it);
• cache-result! ignores a completion whose generation is no longer the node's (an older run finishing after a newer one entered — a boolean mark could not tell the two runs apart); for the current generation it stores the result but certifies it :clean only if that generation was not invalidated. The mark is not consumed, so a second write of the same run (a reject path re-caching through abort-spin-chain!, which passes the same generation) cannot certify it either; nor can a later track resume of the same run. A node that never entered a run has no generation and is only marked :dirty;
• a stale run is never joined: try-claim-execution! and deref start a fresh run instead. The fresh entry (mark-running!, in the same transaction as the generation stamp) hands the superseded run's never-served invoker callbacks to the new run (:run-callbacks → :pending-callbacks); cache-result! drops the record in the same transaction as the run's completing write. The contract is at most once per world: within a world a caller callback is served by its run or by the run it was handed to, never both. The record is fork-local, so each world decides for itself; a :causal-follow callback is served once per world that completes the run — the engine's existing cross-world fan-out (forking.md), which a savepoint's terminal site dedups per world. Nothing is shared between worlds to arbitrate it (that would leak a fork's completion into its parent's decisions);
• a superseding run that completes inline (the await slow path) still enqueues :spin-completion when other awaiters are subscribed to the child — the superseded run's completion is ignored as outdated, so the event is their only wakeup. The inline awaiter's own cont is claimed inline, so it is not resumed twice.

So whoever reaches the node next — inside the batch or in a later drain — re-runs it against the current values. Nothing schedules that re-run by itself: an awaiter already parked on the stale run receives the stale value, and catches up only if the node is re-run and re-completes (its reactive await cont then fires) — e.g. when its creator reaches it. A stale node nobody reaches again stays :dirty.

Why the descendant must be invalidated eagerly: its freshness used to rest on the batch-mode await fast-path guard (await-spin refuses a cache hit during a batch unless the child was processed in it). That covers an ancestor whose re-run reaches the child synchronously within the batch. An ancestor that suspends first — an ifor-each awaiting a newly prepended item that awaits a deferred — reaches the old child only after the batch has ended; the child's still-:clean cache (or its in-flight run, completing on the delivery) was then served as is, permanently. (Regression: engine/stale_run_test.cljc; the differential DOM fuzzer's defect 1 in dom/differential_fuzz_regressions_test.cljs.) This also makes events-settled?'s promise ("what is clean is current") true for descendant observers.

Superseded children (GC of a removed descendant). An item removed by a change is never re-created, but it keeps its own track continuation on the signal — and that continuation reaches the item's Spin object (body continuation → the resolve its awaiter passed in → the awaiter's loop state, e.g. an ifor-each holding all its item spins), so the Spin is never collected and the Cleaner never runs. gc-quiescent? would refuse it anyway (track subscription), and the next change of the signal resumed it as an independent observer. The rule now:

• A run of a creator holds a set of children, live-children: those it (re-)registered (:created-spins) plus, after a track resume, those created before the resume point (:kept-children, snapshotted on the track cont as :slice-state :live-children), which the resumed slice does not re-evaluate. Each new run records what it starts out not holding (:dropped-children).
• When the creator's run completes (record-deps! → release-superseded-children!), ONE transaction (detach-superseded) re-checks and detaches every dropped child that still names that creator, is not held by it, is observed by no other spin, has no run in flight and no caller waiting on one, and is owned by this world — then, transitively, what each detached child held. Detaching drops its conts and subscriptions (one pass over :subscriptions for the whole set), prunes its observer edges and empties its deps, and sets :detached?. Only a child that could have gone stale (signal-dependent?: it tracked a signal, directly or through a spin it awaited — an await alone leaves a persistent :await-reactive cont and does not count) is marked dirty; any other keeps its clean result, so a holder of its Spin awaiting it later does not re-run its effects. A child whose run was in flight is checked when that run completes (error completions too).
• Not while the context has a reachable following fork (ec/following-forks?): a following fork reads every node it has not written through from this world, so a detach decided by this world's creator would also detach the child in a fork whose own creator still holds it. The candidates stay recorded and the generation sweep retries once the fork is gone. The release, the sweep's retry and the deferral record only touch nodes the running world owns: a fork's sweep never acts on a creator it reads through from its parent (writing even its :dropped-children would copy it into the overlay and stop the fork following it). On a JS runtime without WeakRef a fork cannot be tracked without pinning it, so the registry records an unreferenced marker that never expires: such a context defers detaching for good once it has had a following fork.
• A detached node is quiescent even if it was invalidated and never re-ran (gc-quiescent? accepts :detached? for :completed?): it is reaped once its Spin is collected, or in the same transaction if it already was. superseded-child? makes gc-quiescent? accept an orphan whose creator exists, is complete and no longer holds it, despite remaining track conts; a creator absent from the world proves nothing (*spin-id* can name a spin of another context).

Regressions: engine/context_retention_test.clj, and the jsdom differential removed-item-is-reaped-after-same-change-removal.

Open subtlety. Conts with :order less than the resumed one are kept and their signal deps re-tracked (engine/impl/simple.cljc:357-361) — but truncate-stale-conts! re-tracks via track-signal-dep!, which does not refresh :consumed-generation. See 1.7 — flagged.

3.5 Sibling spins

Invariant. Two spins observing the same signal both re-run on a change, in topological order, each at most once.

Argument. ordered-observers returns both in a Kahn topological order (engine/impl/graph.cljc:63-81); the JVM path may resume independent observers in parallel on the executor with a CountDownLatch join (engine/impl/simple.cljc:603-623) — safe because each resume is atomic swap-state! on its own node. CLJS is always sequential. Holds.

3.6 parallel / race

Invariant (parallel). Completes when all children complete; result is the vector of child results; reactive — re-fires when a child re-completes.

Argument. parallel is a :resource spin (spin/combinators.cljc:66-163). It registers an :await-reactive cont per child (combinators.cljc:111-141) so a child's later re-completion re-caches parallel's result and enqueues a fresh :spin-completion. Fail-fast: first child error cancels siblings (combinators.cljc:145-151). Holds. Fragile: coordination atoms completed/done? are local closures, not in context state — the docstring admits they are "write-once, not forked" (combinators.cljc:63-64). A fork of a context with an in-flight parallel shares those atoms — correct only because they are write-once.

Invariant (race). Completes with the first child; losers cancelled. Argument. compare-and-set! on done? picks the winner (combinators.cljc:246-265); cancellation errors from losers are swallowed. Holds.

3.7 debounce / throttle / sample

Invariant. Rate-control combinators delay / coalesce value delivery without losing the final value.

Argument. Each is a :computation spin composing a :resource sleep then an await of the source (combinators.cljc:300-353). When the source signal changes, the computation spin re-runs from the top (it tracks the source transitively), restarting the sleep — that is the debounce timer reset. Holds for value delivery. Scope: the namespace docstrings note rate control "may lose intermediate deltas" unless wrapped in accumulate (combinators.cljc:455-456) — i.e. the delta history is not preserved by debounce/throttle alone, only the snapshot. accumulate (combinators.cljc:457-466) uses a closure atom that persists across re-runs to merge intervals; its correctness depends on merge-fn being associative (a stated precondition, combinators.cljc:425-427).

3.8 Deep nesting under await-cascade

Invariant. A → awaits → B → awaits → C: when C completes, the completion propagates up through B to A, each parent resuming once.

Argument. Each completion enqueues its own :spin-completion; process-event! resumes the immediate parents, whose own resolution enqueues the next :spin-completion — a cascade through the FIFO queue. propagate-await-dirty! (engine/impl/simple.cljc:1658-1711) marks completed, non-suspended ancestors dirty so a re-run reaches stale caches. Holds. Fragile: find-root-await-ancestor (engine/impl/simple.cljc:463-490) walks :observers taking (first parents) — for a spin awaited by multiple parents it picks an arbitrary one as "the" root. Escalation to one root is fine when the await graph is a tree; for a DAG with shared awaited children the "root" is non-deterministic. Flagged as a latent question.

3.9 Rendering identity (delta-direct DOM)

Rendering has no tree diff. build-element reconciles each element's attrs and children against caches keyed by its address, attaches the resulting :deltas to the vnode, and update-render! discharges those deltas against the DOM element bound to that address. Two consequences follow, and both were violated in practice before they were stated:

R1 — Identity is the address. reconcile-slot decides "did this child slot change?" and discharge-vnode! decides "which DOM node do I write to." They must agree, and the only thing the DOM layer binds by is :addr. A vnode's address is content-hash[source-loc, parent-addr, slot-index], and a keyed element's is keyed-child-address(my-addr, key) — so addr equality implies key equality and structural position, but not conversely. :key on a plain element is a disambiguator for siblings at one source location, never an identity token; position-independent identity is ifor-each's job, and its items travel through the :keyed slot type. Letting :key override :addr in reconciliation makes the parent declare a child "unchanged" while the child's deltas address a node that was never created: the update vanishes with no error. (Regression: dom/branch_flip_test.clj.)

R2 — The root is unaddressable from above. Every node except the mounted root is reachable through some parent's slot reconciliation. The root has no parent, so a spin whose body returns a different root element (a conditional at the top of the spin) cannot be updated by deltas at all — the new root's address has no bound element. update-render! detects a changed root identity and re-mounts, running the full teardown protocol first (call-refs-on-unmount! → render-initial! → flush-pending-evictions!) so foreign nodes release resources and dead caches are evicted, while addresses the new tree re-claims are spared by (pending − live).

Corollary (unmount completeness). Every address the render path writes must be reachable from the teardown walk. flatten-slot splices a :keyed slot's items into :children and drops the KeyedFragment itself, so the ifor-each call-site address — which holds :by-key, i.e. every rendered vnode and its handler closures — is on no vnode. evict-cache! therefore cascades through the evictee's slot cache into its keyed slots. Without that cascade, unmounting any subtree containing an ifor-each leaked the entire rendered list for the lifetime of the context.

Because a changed root re-mounts (and thus loses DOM state), keep a spin's root element structurally stable and branch inside it. This is the loading-state idiom: one container with a stable key, cond/if on the content within.

R3 — Reconciliation happens AT the commit point; the caches advance in the same step as the DOM write. Builds are pure values: build-element and for-each* classify slots and package items — no cache reads, no diffing, no staged deltas. At the commit points (initial-mount! seeds caches from the mounted tree; update-render! runs commit-reconcile!, dom/commit) the arrived tree is diffed against the committed per-address caches and applied, and the caches are advanced in one step.

Why this is a rule and not an implementation detail: it is what makes both abandonment AND repetition free. An abandoned body slice never touched anything, so nothing needs undoing. A body executed N times for one logical change — routine when a parent's :await-reactive cont re-fires per child re-completion — produces N pure values; the first commit advances the baseline, and every later build diffs against it and collapses to its genuine residual. The earlier design (compute-time reconciliation with staged caches promoted at commit) could not have the second property: N racing builds all diffed one uncommitted baseline and each carried the same :add, more than one of which could reach the DOM in separate passes (measured live: six builds of one container per settled expand, the same :add discharged twice); and the shared last-write-wins staging slot meant the committed baseline could advance past what the DOM actually held. (Acceptance: dom/commit_reconcile_test.clj — five distinct builds, one insert; dom/jsdom_final_state_test.cljs — the app-shape protocol on a real tree.)

One prerequisite the rule imposes on builds: the vnode carries its slot structure (:slots), because flatten-slots destroys the :nil-slot information the commit diff needs for stable slot indices. And the diff runs against the ARRIVED TREE only — never against any build-side accumulation.

R4 — The environment travels with the continuation, never with the event. A body slice always executes in the environment captured at its own suspension point (or construction, for a first slice): (:bindings ctx), the addressing chain-head, and the transient dep tracking. The environment of the COMPLETER — whoever delivered the value that woke the slice — must never be observable. Enforced at three places: (1) restore-slice-state! REPLACES the context's bindings with the cont's snapshot (merge restored presence but not absence — a key the slice never bound leaked in from the resume context); (2) every continuation carries :slice-state, including :external-await conts (Deferred/Mailbox/thunk), whose :cont-resume and reject paths restore it; (3) the drain NORMALIZES its context to the creation-time :base-bindings — enqueue closures smuggle body-scoped context values into trigger-drain!, and an un-normalized drain executed whole cascades (including FIRST slices of fresh spins, whose construction scope is a delta merged onto the drain context) in a random body's environment. Violation symptom, measured before the rule: an element built after (await (for-each* …)) was addressed under the last-completing ITEM's keyed scope, so its address changed with completion order — R1 broken nondeterministically, the subtree re-minted every render. (Regression: dom/jsdom_final_state_test.cljs fragment-parent-keeps-its-identity-across-rerender; full derivation in .internal/slice-environment-integrity.md.)

R5 — Mount refs are a commit-time effect. render-initial! builds a subtree DETACHED; the parent's append-child! / insert-child! / replace-child! (or the container's appendChild) attaches it only after it returns. A :ref fired from inside it — where it used to fire — saw a node with no parent: .isConnected false, .closest nil, zero layout; every consumer needing an ancestor or a measurement deferred with requestAnimationFrame. Every commit point (initial-mount!, the R2 root re-mount, commit-update! / commit-reconcile!, browser/mount!) now runs inside disch/with-mount-effects: render-initial! QUEUES mount refs, and the queue is flushed once the commit step has finished — all DOM writes landed, caches seeded/advanced, dead addresses retired. Contract:

• Attached. A mount ref receives an element that is in the committed tree, hence document-connected whenever the render container is.
• Order. Each created subtree contributes its refs in post-order — descendants before their ancestor, siblings left to right (the order render-initial! completes elements in); subtrees contribute in the order the commit walk created them.
• Committed builds only. Builds are pure values (R3), so a build abandoned or superseded before commit never reaches render-initial! and fires nothing; a commit that throws drops its queue.
• Unmounts are not deferred. call-refs-on-unmount! fires (ref nil) during the walk, while the departing element is still attached.

Outside with-mount-effects (direct render-initial! calls in tests and tooling) refs fire inline on the detached subtree — the caller owns attachment there. (Regression: dom/ref_linkage_test.cljs — initial mount, ifor-each insert, conditional flip, root re-mount, superseded build — asserting .isConnected and .closest from inside the callback; dom/ref_test.cljc test-mount-refs-are-a-commit-effect for the queue itself.)

Invariant summary table

composition	glitch-free	no double-exec	no stale-cache	no orphan cont	deterministic id
nested spins	✓	✓ (B-gate)	✓ (B-gate)	✓	✓
track in re-run	✓	✓	✓	✓ (post-fix)	✓
await-cascade	✓ (sync)	✓ (dedup)	✓ (propagate)	✓	✓
track + await	✓	✓ (cancel gate)	⚠ gen-staleness	✓ (truncate)	✓
sibling spins	✓	✓	✓	✓	✓
parallel / race	✓	✓ (CAS)	✓	✓	✗ (resource)
debounce/throttle	✓	✓	⚠ deltas lost	✓	✓ (comp) / ✗ (sleep)
deep await nesting	✓ (sync)	✓	✓	✓	⚠ multi-parent root