docs(sweeps): the sweep reference says how many sessions a sweep opens on a remote Gurobi

CHANGELOG.md

@@ -7,6 +7,8 @@ it releases that version ([RELEASING.md](https://github.com/fluxopt/specsolve/bl
 
 ## Upcoming version
 
+- docs(sweeps): the sweep reference says how many sessions a sweep opens on a remote Gurobi ([#1771](https://github.com/fluxopt/specsolve/pull/1771))
+
 ## 0.1.0 (2026-09-25)
 
 🎉 specsolve is officially on PyPI!

docs/reference/sweeps.md

@@ -243,7 +243,7 @@ than the table. Only `scan_parquet` is a reference.
 | | |
 |---|---|
 | **a partition is a filter on the sources** | Not a narrower index: the containment check refuses parameter rows outside the declared coordinates, so the axis rewrites the rows and the index they are over in one mapping. |
-| **one model, updated per slice** | A serial sweep builds once and [updates](api.md#re-solving-with-new-numbers), and a slice whose structure matches the last keeps the loaded solver. A sweep under `executor=` builds per slice, because a built model does not cross a process. |
+| **one model, updated per slice** | A serial sweep builds once and [updates](api.md#re-solving-with-new-numbers), and a slice whose structure matches the last keeps the loaded solver. A sweep under `executor=` builds per slice, because a built model does not cross a process. The loaded solver holds Gurobi's environment, so a [remote Gurobi](api.md#choosing-a-solver), such as Instant Cloud, opens one session for a serial sweep unless `keep='nothing'`. It opens a new one at each slice that `sweep.metrics` marks `loaded`, and at every slice under `executor=`. |
 | **`keep=` reaches every slice, and the fold chooses none of them** | It defaults to `'solver'`, as [`solve`](api.md#how-much-of-the-session-a-solve-keeps) does. `keep='progress'` has something to carry, since consecutive slices differ by one step; whether that pays is a fact about the *model*. Under `executor=` every slice is a first solve and keeps `'nothing'`. |
 | **the model is asked before it is sliced** | The plan says what each axis can bear ([`separability`](https://mathspec.readthedocs.io/en/latest/reference/reading/#asking-whether-an-axis-can-be-cut)). `EachWindow` needs `into` *windowable*: it refuses a coupling by naming the declaration and the change that would lift it, reads an offset the data decides off the data, requires `length - step` to cover what the rows read ahead, and warns where a `position()` restarts per window. What the rows read *behind* is the rolling-horizon seed, met by the edge policy, and is not refused. `EachCoordinate` is not asked: the spec must not declare the column it slices, so the model never sees the axis. |
 | **the spec is parsed once** | `solve_over` validates it up front, so a spec outside the language fails before the data is touched. Every worker is handed the lowered [program](glossary.md#the-chain), in this process or across one, and none reads the YAML or lowers it again. |