@woosh/meep-engine 3.1.0 → 3.2.0

package/src/engine/physics/INTEPOLATION_SYSTEM_PLAN.md

@@ -1,287 +1,287 @@
-# Interpolation system — build plan
-
-Cross-cutting plan for fixed-step → render interpolation. Spans **render**,
-**physics** (a producer), and **network** (a producer + the existing primitives
-this builds on). Referenced from `engine/physics/PLAN.md`.
-
-> Filename note: created as `INTEPOLATION_SYSTEM_PLAN.md` per the request; rename
-> to `INTERPOLATION_SYSTEM_PLAN.md` if the spelling was unintended.
-
----
-
-## Goal
-
-The simulation advances on a fixed step (`EntityManager.fixedUpdateStepSize`,
-~60 Hz). The renderer runs at a variable, usually higher, display rate on its own
-`requestAnimationFrame` loop. Between two fixed steps the rendered pose is frozen
-at the last step, so motion stutters / temporally aliases — worst at low fixed
-rates or high refresh. This system smooths the discrete per-tick poses across the
-render rate.
-
-## Hard constraints (these shaped every decision)
-
-1. **Write `Transform`, not the renderable.** Scene culling, mesh BVHs, particle
-   systems, trails, and attachment hierarchies are all driven off the ECS
-   `Transform`. Writing an interpolated pose into a three.js object directly
-   would smooth *some* visuals and desync the rest. The interpolated pose must
-   land in `Transform` so the whole engine sees it uniformly. (Bonus: attachment
-   children of an interpolated physics parent then track it **for free**, because
-   the attachment system is push-based on the parent's `Transform.onChanged`.)
-2. **Only interpolate what moved.** Per-frame cost must be O(moving bodies),
-   never O(total). The engine targets massive worlds — tens of thousands of
-   dynamic bodies with only a tiny awake/moving fraction at any moment.
-3. **Don't fight physics.** Physics must never integrate from an interpolated
-   pose. The simulation reads authoritative state; interpolation is a downstream
-   render channel.
-4. **Co-designed with network.** The network package can change (not yet shipped).
-   Interpolation must share machinery with replication, not duplicate it.
-5. **Rewind-safe.** The network can rewind and re-simulate. If a rewind+resim
-   produces the same state, the observer must see the **same** interpolated
-   result — no visual drift from re-running history.
-6. **Agnostic to what it interpolates.** Translation+rotation is one case;
-   network may interpolate other replicated quantities. The mechanism blends
-   *encoded snapshots of any registered component type*.
-
-## Key realization — reuse, don't rebuild
-
-The network package already contains the agnostic interpolation engine. Physics
-becomes a *producer* into it; we do not build a parallel system.
-
-- `network/sim/BinaryInterpolationAdapter.js` —
-  `interpolate(out, source, offA, offB, t)`: type-agnostic blend of two encoded
-  snapshots. Concrete `Transform` / `Vector3` / `Quaternion` adapters exist.
-  **This is constraint 6, already solved.**
-- `network/sim/InterpolationLog.js` — a **tick-indexed ring** (`Map<tick, …>`)
-  of per-component snapshots, recording **only components that changed that tick**
-  (the package's O(mutations)-not-O(entities) bet == constraint 2). Keyed by a
-  plain integer; `interpolate(out, key, type_id, tickA, tickB, t, adapter)`
-  blends two ticks into an out-buffer that is then deserialized into the live
-  component.
-- `network/sim/SimActionExecutor.js` — `before_execute` hook, JSDoc:
-  *"NetworkSession undoes render-time interpolation on remote-owned components"*
-  before reading authoritative state. **The architecture already committed to
-  interpolation writing the live component and the sim restoring authoritative
-  before it reads** (constraints 1 + 3).
-- `network/sim/ActionLog.js` + `network/sim/RewindEngine.js` — rewind restores
-  prior authoritative state; resim replays forward via `onReplay`
-  (constraint 5).
-
-## Locked decisions
-
-1. **Neutral shared primitive.** Lift `InterpolationLog`,
-   `BinaryInterpolationAdapter`, and the blend adapters out of `network/` into a
-   neutral `engine/interpolation/`. Physics and network both produce into it; one
-   render-time pass blends for both. Replication + rewind layer on top **only**
-   for networked entities — single-player physics pays no action-log cost.
-2. **Restore-from-log at step start.** Before each physics tick, reset each moving
-   body's live `Transform` to its latest authoritative tick snapshot (a *quiet*
-   write that bypasses `onChanged`). The solver / narrowphase / broadphase /
-   integrators keep reading & writing `Transform` exactly as today — **no rewiring
-   of the physics hot paths.** The authoritative "cache" is the log's newest tick.
-
-## Glossary (load-bearing terms)
-
-- **Tick** — one fixed simulation step, identified by a monotonic integer *tick
-  number*. The orchestrator's frame number when networked; an internal counter
-  otherwise.
-- **Authoritative pose / endpoint** — the exact pose the simulation produced at
-  the *end* of a tick. The thing physics owns and rewind restores.
-- **Snapshot** — an authoritative pose encoded into the log under
-  `(log key, type_id, tick)` via a serialization adapter.
-- **Log key** — the integer identity under which an entity's snapshots are
-  recorded and looked up in the shared log. `network_id` when the entity is
-  networked (so a server's recording and a client's receipt line up under one
-  identity); a locally-allocated interp id otherwise. Carried on the
-  `Interpolated` component. (See §"Q&A".)
-- **alpha** — `accumulator / fixedUpdateStepSize ∈ [0, 1)`: the fractional
-  progress between the two most-recent *completed* ticks. The blend parameter `t`.
-  We render **between** the last two ticks (one step in the past), so both
-  endpoints always exist — no extrapolation. (See §"Q&A".)
-- **`InterpolationSystem`** — the system that owns the render-time blend pass. It
-  runs in `System.update(dt)` (the variable-rate, rAF-driven cycle), NOT ad-hoc in
-  a `preRender` callback. `update` can fire on ticks where no render happens —
-  harmless (a blend with no following draw is just discarded). If display-exact
-  timing is ever needed, the system can additionally hook `preRender` from its
-  `startup`; `update` is the default.
-- **`Interpolated` component** — opt-in marker physics & render cooperate around.
-  Holds the **log key** (defaults to `-1`; `InterpolationSystem.link` assigns it
-  `entity_id` when still `-1`, so the common case needs no manual setup — the
-  network system can instead set it to the `network_id` directly for replicated
-  entities), the blend type/adapter id, and a `snap` flag. A body without it keeps
-  today's direct-write behavior (this is how physics "sees" who to cooperate
-  with).
-
-## Architecture — one engine, two producers
-
-**Per fixed tick, for bodies carrying `Interpolated`:**
-
-1. **Restore (step start):** reset the live `Transform` to the log's latest
-   authoritative tick — a quiet write (no `onChanged`). This is `before_execute`
-   / "undo render interpolation" applied to physics. Physics then reads truth.
-2. **Step:** physics integrates as today → `Transform` = new authoritative pose.
-3. **Record (step end):** if it actually moved (`|Δ| > ε`), snapshot the pose into
-   the current tick. Sparse by construction → O(moving).
-
-**Per `InterpolationSystem.update(dt)` (render/rAF cadence):** for each
-`Interpolated` entity, pick `(tickA, tickB, t)` and blend → deserialize into the
-live `Transform` (this write *does* fire `onChanged` → culling/particles/trails/
-attachments update). O(moving). `t` source is per-producer: local =
-`EntityManager.getFixedStepAlpha(physicsSystem)`; remote = the network's existing
-render-delay/jitter playout.
-
-```
-simulate()  [fixed-step catch-up loop]      InterpolationSystem.update(dt)  [rAF cadence]
-  for awake+Interpolated:                     for Interpolated e:
-    restore Transform ← log.latest (quiet)      (a,b,t) = window(e)   // t = getFixedStepAlpha(physics)
-  ...physics step → Transform = authoritative   blend(a,b,t) → Transform   (onChanged)
-  for moved+Interpolated:                       → later draw passes read Transform
-    log.record(key, tick, Transform)
-```
-
-## Why each constraint holds
-
-1. **Writes Transform** — blend deserializes into the live component; every
-   Transform-linked system works, hierarchy included.
-2. **Only moving** — produce & blend only for moved bodies; the log is sparse.
-3. **Doesn't fight physics** — restore-at-step-start; physics never sees the
-   interpolated pose. Provable by an interp-on vs interp-off sim-state bit-identity
-   test.
-4. **Co-designed with network** — same log, same adapters, same `before_execute`;
-   physics is a new producer into the existing pipeline.
-5. **Rewind-safe** — the log is tick-keyed; resim re-runs the deterministic sim,
-   which re-records the *same* tick numbers with identical bytes → identical
-   blend output. Requirement: the replay path records snapshots like the live path
-   (it runs the same systems), and the log supports re-recording rewound ticks.
-6. **Agnostic** — `BinaryInterpolationAdapter`; pose is one type; any component
-   with a (serialization adapter + interpolation adapter) pair interpolates the
-   same way.
-
-## Invariants
-
-- **Interpolation never feeds back into the sim.** Guarded by restore-at-step-start
-  and an interp-on/off sim bit-identity test.
-- **The log's latest tick is the single authoritative cache** between steps;
-  `Transform` is a derived render channel for interpolated bodies.
-- **Sparse by motion** — O(moving), never O(N).
-
-## Precision note (massive worlds)
-
-The network *wire* Transform adapter uses Float32 position + smallest-three
-quaternion for bandwidth. That quantizes badly far from the origin. The shared
-**interp log stores full-precision snapshots** (Float64 position + full
-quaternion) for both local and remote (remote is re-serialized at full precision
-on apply); wire compression stays on the wire only.
-
-## Contract change (interpolated bodies only)
-
-For a body carrying `Interpolated`, raw external `Transform` writes are no longer
-authoritative — restore-at-step-start would clobber them. External authoritative
-changes (teleport, kinematic drive) go through a `setPose` / `teleport` API that
-updates the log's latest tick **and** sets `snap` (so the next render does not
-slide across the jump). Bodies *without* `Interpolated` keep today's raw-write
-teleport contract.
-
-## Phases
-
-Each phase: implement → spec → run from `H:/git/moh` → commit. Black-box tests,
-headless where possible (the log, adapters, and physics producer are
-headless-testable; the render hook is thin).
-
-> **Status (Phases 0–2 landed).** Key deltas from the original sketch, decided
-> during build: the fixed-step **tick lives on `EntityManager`**
-> (`fixedStepTick`), not a per-system clock, and `getFixedStepAlpha()` is the one
-> global sim-clock alpha (no system arg). To make the per-step tick exact, the
-> scheduler was changed to a **global lock-step loop** on a single accumulator
-> (commit `0fd41f3`) — all systems' `fixedUpdate` advance together, tick++ per
-> step. The physics producer is **agnostic** (iterates the component's
-> interpolands, uses their adapters) and reads its log via an injected
-> `PhysicsSystem.interpolationLog` (null → skipped). Restore reads `tick-1`,
-> record writes `tick`; lock-step makes both exact with no per-body tick field.
-
-- [x] **Phase 0 — Neutralize the primitive** (`0980e45`). `InterpolationLog` +
-  `BinaryInterpolationAdapter` (+ spec) moved to `engine/interpolation/`; network
-  importers repointed. `EntityManager.getFixedStepAlpha`. (Log key generalization
-  is cosmetic — it was already a plain integer.)
-- [x] **Phase 1 — `Interpolated` + `InterpolationSystem`** (`fc3c3aa`).
-  `Interpoland` descriptor; `Interpolated` carries a **list** of interpolands
-  (one component-per-class → many smoothed quantities); `link` auto-assigns
-  `key = entity` when `-1`; `update` blends each interpoland `(tick-1, tick)` at
-  `getFixedStepAlpha()` into the live component. Full-precision (Float64) pose
-  adapters + `POSE_INTERPOLAND`. Tested: blend / snap / untouched-when-empty /
-  key auto-assign / multi-interpoland.
-- [x] **Phase 2 — Physics as producer** (`a05db76` tick + `dade4d6` producer;
-  scheduler `0fd41f3`). `__interp_restore` (top of `fixedUpdate`, undo render
-  interpolation) + `__interp_record` (after solve/CCD, before sleep test).
-  **Determinism guard passes**: interp ON (live Transform polluted every awake
-  step) vs OFF → bit-identical pose AND velocity every step through fall →
-  contact → settle → sleep.
-  - [x] *`setPose` teleport API* (`691f9a0`). `PhysicsSystem.setPose(rb, pos, rot)`
-    writes the Transform, wakes, and sets `snap`. `snap` semantics finalized:
-    the consumer **leaves the live pose untouched** for that frame (a teleport's
-    new pose isn't in the log yet) and the producer's restore **skips** a snapped
-    body. *Known edge:* clean for the common ~1:1 sim:render rate; when render
-    rate ≫ sim rate a teleport could show a brief flicker (the latest tick isn't
-    retroactively overwritten) — fixable later with an in-place latest-tick
-    overwrite if it ever matters.
-  - [x] *End-to-end wiring verified* (`691f9a0`,
-    `InterpolationSystem.integration.spec`). A real `EntityManager` driving
-    physics + `InterpolationSystem` with `physics.interpolationLog = interp.log`
-    interpolates a body's `Transform` every render frame between fixed steps
-    (control with interp off is frozen between steps). *App wiring* (registering
-    the systems + connecting the log in the actual game world) is still the app
-    layer's job — the recipe is the integration spec.
-- **Phase 3 — Network unification.** *Architecture + coherence landed; live
-  NetworkSession migration is the remaining integration.*
-  - **Finding:** the network package *already* does remote interpolation on the
-    shared primitive — `NetworkSession` owns an `InterpolationLog`, records
-    remote-owned components on `onFrameApplied` (`#on_frame_applied`), undoes
-    render interpolation via `normalize_if_dirty` (the `before_execute` hook),
-    and blends each remote entity through the SAME `log.interpolate → deserialize`
-    path as the local system, on an `AdaptiveRenderDelay`-smoothed playout frame
-    (`#render_interpolated_entities`). So the *primitive* was unified by Phase 0;
-    what was missing was a single *consumer* spanning both timelines.
-  - [x] **Unified consumer.** Component stays pure data: `Interpolated.sourceId`
-    (an int; `INTERPOLATION_SOURCE_LOCAL = 0` default = local fixed-step clock +
-    the system's own log). The timelines' logs + window providers live on the
-    SYSTEM, registered via `InterpolationSystem.registerSource(id, log, sample)`
-    (the network owns its playout window). One `update` blends local and
-    source-driven entities, each source sampled once per frame. *Test:* a local
-    sim-clock entity and a remote playout-source entity both interpolate
-    correctly in one `update` (local + remote coherent in one scene).
-  - [ ] **Live migration (remaining):** point `NetworkSession`'s remote entities
-    at a `NetworkInterpolationSource` (wrapping its playout window + log), mark
-    them with `Interpolated`, and drop `#render_interpolated_entities` so the
-    `InterpolationSystem` is the sole render consumer. Deferred to avoid
-    destabilizing the tested netcode; needs exposing the session's playout window
-    as a side-effect-free `sample()`.
-- **Phase 4 — Rewind/resim correctness.** Interp log supports resim re-recording:
-  on rewind to target, truncate ticks ≥ target; resim re-records forward.
-  *Test:* sim to T, capture interp-log bytes; rewind to T−k, resim to T with
-  identical inputs; assert interp-log bytes over (T−k, T] are bit-identical →
-  observer-identical when nothing changed.
-- **Phase 5 — Camera + edges.** Camera following an interpolated body reads the
-  interpolated source pose. Spawn/despawn & key reuse via generation; verify a
-  visual child of an interpolated physics parent tracks it (hierarchy-for-free);
-  kinematic/parked filtered by the moved-ε test.
-
-## Open items (flagged, not yet decided)
-
-- **Local-vs-remote render timeline.** Local predicted-present vs remote
-  delayed-playout is the classic netcode coherence question. v1 can render both
-  at local-present; refine when the predict/reconcile client lands.
-- **`onChanged` suppression at sim time.** Interpolated bodies could suppress
-  `integrate_position`'s `onChanged` (the render blend will set the final pose).
-  Keep it for now — it keeps culling current between slow render frames.
-
-## Source touchpoints (verify at implementation)
-
-- `engine/physics/integration/integrate_position.js` — writes `Transform` via
-  `.set()` (fires `onChanged`); restore must be a quiet sibling write.
-- `engine/physics/body/BodyStorage.js` — awake set (`awake_at`/`awake_count`) =
-  moving candidates; generation = spawn/despawn/reuse detection.
-- `engine/ecs/EntityManager.js` — `fixedUpdateStepSize`, the per-system
-  accumulator `systemAccumulatedFixedStepTime`, the fixed-step catch-up loop.
-- `engine/graphics/GraphicsEngine.js` — `on.preRender` (the blend hook, fired
-  before the draw passes).
-- `engine/network/sim/{InterpolationLog,BinaryInterpolationAdapter,
-  SimActionExecutor,RewindEngine}.js` — the primitives to neutralize/extend.
+# Interpolation system — build plan
+
+Cross-cutting plan for fixed-step → render interpolation. Spans **render**,
+**physics** (a producer), and **network** (a producer + the existing primitives
+this builds on). Referenced from `engine/physics/PLAN.md`.
+
+> Filename note: created as `INTEPOLATION_SYSTEM_PLAN.md` per the request; rename
+> to `INTERPOLATION_SYSTEM_PLAN.md` if the spelling was unintended.
+
+---
+
+## Goal
+
+The simulation advances on a fixed step (`EntityManager.fixedUpdateStepSize`,
+~60 Hz). The renderer runs at a variable, usually higher, display rate on its own
+`requestAnimationFrame` loop. Between two fixed steps the rendered pose is frozen
+at the last step, so motion stutters / temporally aliases — worst at low fixed
+rates or high refresh. This system smooths the discrete per-tick poses across the
+render rate.
+
+## Hard constraints (these shaped every decision)
+
+1. **Write `Transform`, not the renderable.** Scene culling, mesh BVHs, particle
+   systems, trails, and attachment hierarchies are all driven off the ECS
+   `Transform`. Writing an interpolated pose into a three.js object directly
+   would smooth *some* visuals and desync the rest. The interpolated pose must
+   land in `Transform` so the whole engine sees it uniformly. (Bonus: attachment
+   children of an interpolated physics parent then track it **for free**, because
+   the attachment system is push-based on the parent's `Transform.onChanged`.)
+2. **Only interpolate what moved.** Per-frame cost must be O(moving bodies),
+   never O(total). The engine targets massive worlds — tens of thousands of
+   dynamic bodies with only a tiny awake/moving fraction at any moment.
+3. **Don't fight physics.** Physics must never integrate from an interpolated
+   pose. The simulation reads authoritative state; interpolation is a downstream
+   render channel.
+4. **Co-designed with network.** The network package can change (not yet shipped).
+   Interpolation must share machinery with replication, not duplicate it.
+5. **Rewind-safe.** The network can rewind and re-simulate. If a rewind+resim
+   produces the same state, the observer must see the **same** interpolated
+   result — no visual drift from re-running history.
+6. **Agnostic to what it interpolates.** Translation+rotation is one case;
+   network may interpolate other replicated quantities. The mechanism blends
+   *encoded snapshots of any registered component type*.
+
+## Key realization — reuse, don't rebuild
+
+The network package already contains the agnostic interpolation engine. Physics
+becomes a *producer* into it; we do not build a parallel system.
+
+- `network/sim/BinaryInterpolationAdapter.js` —
+  `interpolate(out, source, offA, offB, t)`: type-agnostic blend of two encoded
+  snapshots. Concrete `Transform` / `Vector3` / `Quaternion` adapters exist.
+  **This is constraint 6, already solved.**
+- `network/sim/InterpolationLog.js` — a **tick-indexed ring** (`Map<tick, …>`)
+  of per-component snapshots, recording **only components that changed that tick**
+  (the package's O(mutations)-not-O(entities) bet == constraint 2). Keyed by a
+  plain integer; `interpolate(out, key, type_id, tickA, tickB, t, adapter)`
+  blends two ticks into an out-buffer that is then deserialized into the live
+  component.
+- `network/sim/SimActionExecutor.js` — `before_execute` hook, JSDoc:
+  *"NetworkSession undoes render-time interpolation on remote-owned components"*
+  before reading authoritative state. **The architecture already committed to
+  interpolation writing the live component and the sim restoring authoritative
+  before it reads** (constraints 1 + 3).
+- `network/sim/ActionLog.js` + `network/sim/RewindEngine.js` — rewind restores
+  prior authoritative state; resim replays forward via `onReplay`
+  (constraint 5).
+
+## Locked decisions
+
+1. **Neutral shared primitive.** Lift `InterpolationLog`,
+   `BinaryInterpolationAdapter`, and the blend adapters out of `network/` into a
+   neutral `engine/interpolation/`. Physics and network both produce into it; one
+   render-time pass blends for both. Replication + rewind layer on top **only**
+   for networked entities — single-player physics pays no action-log cost.
+2. **Restore-from-log at step start.** Before each physics tick, reset each moving
+   body's live `Transform` to its latest authoritative tick snapshot (a *quiet*
+   write that bypasses `onChanged`). The solver / narrowphase / broadphase /
+   integrators keep reading & writing `Transform` exactly as today — **no rewiring
+   of the physics hot paths.** The authoritative "cache" is the log's newest tick.
+
+## Glossary (load-bearing terms)
+
+- **Tick** — one fixed simulation step, identified by a monotonic integer *tick
+  number*. The orchestrator's frame number when networked; an internal counter
+  otherwise.
+- **Authoritative pose / endpoint** — the exact pose the simulation produced at
+  the *end* of a tick. The thing physics owns and rewind restores.
+- **Snapshot** — an authoritative pose encoded into the log under
+  `(log key, type_id, tick)` via a serialization adapter.
+- **Log key** — the integer identity under which an entity's snapshots are
+  recorded and looked up in the shared log. `network_id` when the entity is
+  networked (so a server's recording and a client's receipt line up under one
+  identity); a locally-allocated interp id otherwise. Carried on the
+  `Interpolated` component. (See §"Q&A".)
+- **alpha** — `accumulator / fixedUpdateStepSize ∈ [0, 1)`: the fractional
+  progress between the two most-recent *completed* ticks. The blend parameter `t`.
+  We render **between** the last two ticks (one step in the past), so both
+  endpoints always exist — no extrapolation. (See §"Q&A".)
+- **`InterpolationSystem`** — the system that owns the render-time blend pass. It
+  runs in `System.update(dt)` (the variable-rate, rAF-driven cycle), NOT ad-hoc in
+  a `preRender` callback. `update` can fire on ticks where no render happens —
+  harmless (a blend with no following draw is just discarded). If display-exact
+  timing is ever needed, the system can additionally hook `preRender` from its
+  `startup`; `update` is the default.
+- **`Interpolated` component** — opt-in marker physics & render cooperate around.
+  Holds the **log key** (defaults to `-1`; `InterpolationSystem.link` assigns it
+  `entity_id` when still `-1`, so the common case needs no manual setup — the
+  network system can instead set it to the `network_id` directly for replicated
+  entities), the blend type/adapter id, and a `snap` flag. A body without it keeps
+  today's direct-write behavior (this is how physics "sees" who to cooperate
+  with).
+
+## Architecture — one engine, two producers
+
+**Per fixed tick, for bodies carrying `Interpolated`:**
+
+1. **Restore (step start):** reset the live `Transform` to the log's latest
+   authoritative tick — a quiet write (no `onChanged`). This is `before_execute`
+   / "undo render interpolation" applied to physics. Physics then reads truth.
+2. **Step:** physics integrates as today → `Transform` = new authoritative pose.
+3. **Record (step end):** if it actually moved (`|Δ| > ε`), snapshot the pose into
+   the current tick. Sparse by construction → O(moving).
+
+**Per `InterpolationSystem.update(dt)` (render/rAF cadence):** for each
+`Interpolated` entity, pick `(tickA, tickB, t)` and blend → deserialize into the
+live `Transform` (this write *does* fire `onChanged` → culling/particles/trails/
+attachments update). O(moving). `t` source is per-producer: local =
+`EntityManager.getFixedStepAlpha(physicsSystem)`; remote = the network's existing
+render-delay/jitter playout.
+
+```
+simulate()  [fixed-step catch-up loop]      InterpolationSystem.update(dt)  [rAF cadence]
+  for awake+Interpolated:                     for Interpolated e:
+    restore Transform ← log.latest (quiet)      (a,b,t) = window(e)   // t = getFixedStepAlpha(physics)
+  ...physics step → Transform = authoritative   blend(a,b,t) → Transform   (onChanged)
+  for moved+Interpolated:                       → later draw passes read Transform
+    log.record(key, tick, Transform)
+```
+
+## Why each constraint holds
+
+1. **Writes Transform** — blend deserializes into the live component; every
+   Transform-linked system works, hierarchy included.
+2. **Only moving** — produce & blend only for moved bodies; the log is sparse.
+3. **Doesn't fight physics** — restore-at-step-start; physics never sees the
+   interpolated pose. Provable by an interp-on vs interp-off sim-state bit-identity
+   test.
+4. **Co-designed with network** — same log, same adapters, same `before_execute`;
+   physics is a new producer into the existing pipeline.
+5. **Rewind-safe** — the log is tick-keyed; resim re-runs the deterministic sim,
+   which re-records the *same* tick numbers with identical bytes → identical
+   blend output. Requirement: the replay path records snapshots like the live path
+   (it runs the same systems), and the log supports re-recording rewound ticks.
+6. **Agnostic** — `BinaryInterpolationAdapter`; pose is one type; any component
+   with a (serialization adapter + interpolation adapter) pair interpolates the
+   same way.
+
+## Invariants
+
+- **Interpolation never feeds back into the sim.** Guarded by restore-at-step-start
+  and an interp-on/off sim bit-identity test.
+- **The log's latest tick is the single authoritative cache** between steps;
+  `Transform` is a derived render channel for interpolated bodies.
+- **Sparse by motion** — O(moving), never O(N).
+
+## Precision note (massive worlds)
+
+The network *wire* Transform adapter uses Float32 position + smallest-three
+quaternion for bandwidth. That quantizes badly far from the origin. The shared
+**interp log stores full-precision snapshots** (Float64 position + full
+quaternion) for both local and remote (remote is re-serialized at full precision
+on apply); wire compression stays on the wire only.
+
+## Contract change (interpolated bodies only)
+
+For a body carrying `Interpolated`, raw external `Transform` writes are no longer
+authoritative — restore-at-step-start would clobber them. External authoritative
+changes (teleport, kinematic drive) go through a `setPose` / `teleport` API that
+updates the log's latest tick **and** sets `snap` (so the next render does not
+slide across the jump). Bodies *without* `Interpolated` keep today's raw-write
+teleport contract.
+
+## Phases
+
+Each phase: implement → spec → run from `H:/git/moh` → commit. Black-box tests,
+headless where possible (the log, adapters, and physics producer are
+headless-testable; the render hook is thin).
+
+> **Status (Phases 0–2 landed).** Key deltas from the original sketch, decided
+> during build: the fixed-step **tick lives on `EntityManager`**
+> (`fixedStepTick`), not a per-system clock, and `getFixedStepAlpha()` is the one
+> global sim-clock alpha (no system arg). To make the per-step tick exact, the
+> scheduler was changed to a **global lock-step loop** on a single accumulator
+> (commit `0fd41f3`) — all systems' `fixedUpdate` advance together, tick++ per
+> step. The physics producer is **agnostic** (iterates the component's
+> interpolands, uses their adapters) and reads its log via an injected
+> `PhysicsSystem.interpolationLog` (null → skipped). Restore reads `tick-1`,
+> record writes `tick`; lock-step makes both exact with no per-body tick field.
+
+- [x] **Phase 0 — Neutralize the primitive** (`0980e45`). `InterpolationLog` +
+  `BinaryInterpolationAdapter` (+ spec) moved to `engine/interpolation/`; network
+  importers repointed. `EntityManager.getFixedStepAlpha`. (Log key generalization
+  is cosmetic — it was already a plain integer.)
+- [x] **Phase 1 — `Interpolated` + `InterpolationSystem`** (`fc3c3aa`).
+  `Interpoland` descriptor; `Interpolated` carries a **list** of interpolands
+  (one component-per-class → many smoothed quantities); `link` auto-assigns
+  `key = entity` when `-1`; `update` blends each interpoland `(tick-1, tick)` at
+  `getFixedStepAlpha()` into the live component. Full-precision (Float64) pose
+  adapters + `POSE_INTERPOLAND`. Tested: blend / snap / untouched-when-empty /
+  key auto-assign / multi-interpoland.
+- [x] **Phase 2 — Physics as producer** (`a05db76` tick + `dade4d6` producer;
+  scheduler `0fd41f3`). `__interp_restore` (top of `fixedUpdate`, undo render
+  interpolation) + `__interp_record` (after solve/CCD, before sleep test).
+  **Determinism guard passes**: interp ON (live Transform polluted every awake
+  step) vs OFF → bit-identical pose AND velocity every step through fall →
+  contact → settle → sleep.
+  - [x] *`setPose` teleport API* (`691f9a0`). `PhysicsSystem.setPose(rb, pos, rot)`
+    writes the Transform, wakes, and sets `snap`. `snap` semantics finalized:
+    the consumer **leaves the live pose untouched** for that frame (a teleport's
+    new pose isn't in the log yet) and the producer's restore **skips** a snapped
+    body. *Known edge:* clean for the common ~1:1 sim:render rate; when render
+    rate ≫ sim rate a teleport could show a brief flicker (the latest tick isn't
+    retroactively overwritten) — fixable later with an in-place latest-tick
+    overwrite if it ever matters.
+  - [x] *End-to-end wiring verified* (`691f9a0`,
+    `InterpolationSystem.integration.spec`). A real `EntityManager` driving
+    physics + `InterpolationSystem` with `physics.interpolationLog = interp.log`
+    interpolates a body's `Transform` every render frame between fixed steps
+    (control with interp off is frozen between steps). *App wiring* (registering
+    the systems + connecting the log in the actual game world) is still the app
+    layer's job — the recipe is the integration spec.
+- **Phase 3 — Network unification.** *Architecture + coherence landed; live
+  NetworkSession migration is the remaining integration.*
+  - **Finding:** the network package *already* does remote interpolation on the
+    shared primitive — `NetworkSession` owns an `InterpolationLog`, records
+    remote-owned components on `onFrameApplied` (`#on_frame_applied`), undoes
+    render interpolation via `normalize_if_dirty` (the `before_execute` hook),
+    and blends each remote entity through the SAME `log.interpolate → deserialize`
+    path as the local system, on an `AdaptiveRenderDelay`-smoothed playout frame
+    (`#render_interpolated_entities`). So the *primitive* was unified by Phase 0;
+    what was missing was a single *consumer* spanning both timelines.
+  - [x] **Unified consumer.** Component stays pure data: `Interpolated.sourceId`
+    (an int; `INTERPOLATION_SOURCE_LOCAL = 0` default = local fixed-step clock +
+    the system's own log). The timelines' logs + window providers live on the
+    SYSTEM, registered via `InterpolationSystem.registerSource(id, log, sample)`
+    (the network owns its playout window). One `update` blends local and
+    source-driven entities, each source sampled once per frame. *Test:* a local
+    sim-clock entity and a remote playout-source entity both interpolate
+    correctly in one `update` (local + remote coherent in one scene).
+  - [ ] **Live migration (remaining):** point `NetworkSession`'s remote entities
+    at a `NetworkInterpolationSource` (wrapping its playout window + log), mark
+    them with `Interpolated`, and drop `#render_interpolated_entities` so the
+    `InterpolationSystem` is the sole render consumer. Deferred to avoid
+    destabilizing the tested netcode; needs exposing the session's playout window
+    as a side-effect-free `sample()`.
+- **Phase 4 — Rewind/resim correctness.** Interp log supports resim re-recording:
+  on rewind to target, truncate ticks ≥ target; resim re-records forward.
+  *Test:* sim to T, capture interp-log bytes; rewind to T−k, resim to T with
+  identical inputs; assert interp-log bytes over (T−k, T] are bit-identical →
+  observer-identical when nothing changed.
+- **Phase 5 — Camera + edges.** Camera following an interpolated body reads the
+  interpolated source pose. Spawn/despawn & key reuse via generation; verify a
+  visual child of an interpolated physics parent tracks it (hierarchy-for-free);
+  kinematic/parked filtered by the moved-ε test.
+
+## Open items (flagged, not yet decided)
+
+- **Local-vs-remote render timeline.** Local predicted-present vs remote
+  delayed-playout is the classic netcode coherence question. v1 can render both
+  at local-present; refine when the predict/reconcile client lands.
+- **`onChanged` suppression at sim time.** Interpolated bodies could suppress
+  `integrate_position`'s `onChanged` (the render blend will set the final pose).
+  Keep it for now — it keeps culling current between slow render frames.
+
+## Source touchpoints (verify at implementation)
+
+- `engine/physics/integration/integrate_position.js` — writes `Transform` via
+  `.set()` (fires `onChanged`); restore must be a quiet sibling write.
+- `engine/physics/body/BodyStorage.js` — awake set (`awake_at`/`awake_count`) =
+  moving candidates; generation = spawn/despawn/reuse detection.
+- `engine/ecs/EntityManager.js` — `fixedUpdateStepSize`, the per-system
+  accumulator `systemAccumulatedFixedStepTime`, the fixed-step catch-up loop.
+- `engine/graphics/GraphicsEngine.js` — `on.preRender` (the blend hook, fired
+  before the draw passes).
+- `engine/network/sim/{InterpolationLog,BinaryInterpolationAdapter,
+  SimActionExecutor,RewindEngine}.js` — the primitives to neutralize/extend.