docs: update skill guides for auto-tracking and batching changes

.agents/skills/developing-oecs/SKILL.md

@@ -38,8 +38,8 @@ Key ADRs to keep in mind when changing the library:
 | 002 | 32-bit Entity (24 ID + 8 version) | Fits in register, enough headroom |
 | 003 | Explicit query iteration | Visible cost model, no code-gen needed |
 | 004 | Registration order for systems | Simplest model that works at this scale |
-| 005 | Manual `MarkModified` for value changes | No per-component overhead |
-| 006 | Deferred change posting | Prevents mid-system reentrancy |
+| 005 | Auto-tracking in batching scopes | Eliminates manual `MarkModified` in systems |
+| 006 | Deferred mutation batching | Prevents mid-system reentrancy, safe iteration |
 | 007 | R3 for reactivity | Zero-allocation, UI lifecycle-friendly |
 | 008 | Commands as serializable structs in a queue | Not ECS state |
 | 009 | Singleton per component type | Each singleton gets its own entity |
@@ -73,12 +73,12 @@ Uses swap-remove for O(1) deletion.
 
 Zero-allocation `ref struct` enumerators returned by `world.Select<T1..T6>()`.
 Drive from the smallest sparse set to minimize probes. Support `foreach` via
-`GetEnumerator()` returning `this`. Expose `Entity` and `ref Item1`..`ref ItemN`
-properties. Constructor/dispose manage `BeginIteration()`/`EndIteration()` for
-pending mutation flushing. Singleton entities are excluded via
-`IsSingletonEntity()` check.
+`GetEnumerator()` returning `this`. Expose `Entity`, `Ref1`..`RefN` (mutable,
+tracked for auto-dirty-marking), and `Val1`..`ValN` (read-only, untracked).
+Constructor/dispose manage `BeginBatching()`/`EndBatching()` for pending
+mutation flushing. Singleton entities are excluded via `IsSingletonEntity()`.
 
-Also provides `FindEntity<T>()` (first match) and `FindEntities<T>()` (all matches).
+Also provides `FindEntity<T>()` (first match).
 
 ### `Query<T1..T6>` — generic query descriptors
 
@@ -90,8 +90,11 @@ method adds exclusion filters without changing the arity. Replaces the old
 
 Manages an ordered list of `ISystem`. `RunAll()`:
 1. Drain commands (pre-tick).
-2. For each system: run it → drain commands → post changes.
-3. Drain commands + post changes (post-tick).
+2. `BeginBatching()` — outer batching scope for the tick.
+3. For each system: run via `Run` extension (nested batch) → drain commands →
+   flush pending mutations → post changes.
+4. `EndBatching()` — auto-marks + flushes remaining.
+5. Drain commands + post changes (post-tick).
 
 ### `WorldSerializer` — save/load
 
@@ -189,11 +192,12 @@ when there are no fields to compare.
 
 ## Common Pitfalls
 
-- **Forgetting `MarkModified` after mutating a `ref` component.** Changes won't
-  propagate to R3 subscribers. Debug builds have a warning for this.
+- **Using `ValN` when you meant to mutate.** `ValN` returns a `ref readonly` —
+  writes through it still work but won't be tracked for auto-dirty-marking.
+  Use `RefN` for any component you intend to mutate.
 - **Modifying components during iteration.** Pending mutations are flushed at
-  the end of iteration (via `BeginIteration`/`EndIteration`). Adding/removing
-  components mid-iteration is safe — they're deferred.
+  the end of the batching scope. Adding/removing components mid-iteration is
+  safe — they're deferred.
 - **Type not public or missing `[MessagePackObject]`.** Serialization will fail.
   The `MessagePackAnalyzer` catches most issues at compile time.
 - **Singletons each have their own entity allocated automatically by