docs: FIXME — don't advance schema bundle on shortage

During a shortage streak the published worker assignment stays frozen
while the schema bundle keeps advancing. A worker joining mid-streak
downloads the frozen assignment but can't resolve the GC'd schema of a
still-named (removing) chunk — the failure schema_bundle_consistency
checks for, currently masked by its excluded guard.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: define-null

src/multistep_scheduler/sim/sut.rs

@@ -897,6 +897,11 @@ impl<D: SimStorage> SimUnderTest<D> {
                 Some(id)
             }
             Err(StorageError::Shortage) => {
+                // FIXME: on a shortage the published worker assignment stays frozen while the
+                // schema bundle keeps advancing, so a worker joining mid-streak downloads the
+                // frozen assignment and can't resolve the GC'd schema of a still-named (removing)
+                // chunk. Fix: don't advance the schema bundle on `Shortage` — keep it in lockstep
+                // with the worker assignment clients consume. See `schema_bundle_consistency`.
                 self.schedule_status = ScheduleStatus::NotEnoughCapacity;
                 self.is_infeasible = true;
                 None

src/multistep_scheduler/sim/sut/placement_oracles.rs

@@ -127,6 +127,14 @@ pub(super) fn published_coverage(
 /// `worker_assignment` is a cache only refreshed on a successful scheduling cycle, so during a
 /// shortage streak it can lag behind the freshly-regenerated `bundle` — a lagging chunk's schema
 /// may have already been legitimately GC'd from the bundle, which is not a resolution failure.
+///
+/// FIXME: the exclusion masks a real production hazard rather than a benign lag. During a shortage
+/// streak the published worker assignment stays frozen while the bundle keeps advancing, so a
+/// worker that joins mid-streak downloads that frozen assignment and then cannot resolve the
+/// GC'd schema of a still-named (removing) chunk — the exact failure this oracle exists to catch.
+/// Fix: don't advance the schema bundle when a cycle hits `Shortage`; keep it in lockstep with the
+/// worker assignment that clients actually consume, so a schema is never GC'd while an assignment
+/// still references its chunk. Once that holds, this exclusion can drop.
 pub(super) fn schema_bundle_consistency(
     bundle: &SchemaBundle,
     worker_assignment: Option<&WorkerAssignment>,