fix: let a window stop at each group's edge, as its checks already assumed

[FBumann]

Closes #57. Stacked on #56 — base is feat/sum-forward, so review that first.

What was wrong

by= was half-landed. dimensions.py partition-checked sum_back, and operators.py's docstring said by= worked on it — but the Builtin had no lookup_kwargs, so call_shape_error refused the call before either could run. Checked, documented, unreachable.

What was missing

Only the render. Resolution is generic (it reads lookup_kwargs), so the signature change alone makes by= arrive as a LookupNode and the existing dimension checks light up. What had to be built is the group on the operator inside the domain, where shift carries it on a leaf subscript instead:

$$\sum_{h' \in \mathcal{H} ,:, 0 \le h \ominus^{\mathrm{day_of}(h)} h' < 3} \mathit{started}_{u,h'} \le \mathit{on}_{u,h}$$

The superscript takes the bare index, not the subscript in force — the group is a property of the row being written, and a window over a translated operand still asks which group that row is in. That matches _Context.subscript, which does the same for shift.

Why it's needed

Representative periods. A dozen typical days stand in for a year, each weighted, their hours numbered consecutively — but that numbering is storage order, not a timeline. The last hour of one representative day does not precede the first of the next in any physical sense.

An unpartitioned window sums straight across that seam. It builds, it solves, and the commitment it returns couples two independent samples through a boundary that doesn't exist. Nothing warns, because the axis is ordered and the language has no other way to learn the order isn't time.

Not sayable another way, so this is Capability rather than ergonomics under #30: the partition is a condition inside the summation's domain, relating the bound index to the row's. A where masks which rows get built, not how far one reaches — excluding rows near a boundary would drop the constraints entirely rather than shorten the windows.

Notes for review

• The -^{group} typography is pre-existing. A bare partitioned window renders 0 \le h -^{\mathrm{day\_of}(h)} h' < 3, a plain minus carrying a superscript. shift already does exactly this — x_{h -^{\mathrm{day\_of}(h)} 1} — so the windows now follow the established convention rather than inventing one. If that convention is worth revisiting, it should be revisited for both.
• test_the_golden_model_reaches_every_line_of_the_walk caught the missing golden coverage before I did; seasonal_window is there for it.
• Two new probes (sum_back_partitioned, sum_forward_partitioned), two doc-table rows, and a prose section leading with the representative-periods case.
• 438 passing, lint clean.

🤖 Generated with Claude Code

[read-the-docs-community]

Documentation build overview

📚 math-spec | 🛠️ Build #34211206 | 📁 Comparing 22e25ec against latest (f1c207f)

  🔍 Preview build  

6 files changed · ± 6 modified

± Modified

• examples/operators/index.html
• reference/notation/index.html
• reference/language/operators/index.html
• reference/math_spec/dimensions/index.html
• reference/math_spec/operators/index.html
• reference/math_spec/typesetting/walk/index.html

[FBumann]

Split in two and closing as superseded — the change turned out to be two independent ones wearing a single commit.

• fix: let sum_back stop at each group's edge, as its checks already assumed #65 takes the sum_back half, straight to main. That half was not waiting on sum_forward at all: the BUILTINS comment already said by= partitions the axis for "shift and sum_back", and _dims_call already ran every partition check for both names. Only the builtin never declared the kwarg, so the rule two other places described was unreachable. The rendering there is identical to this PR's for seasonal_window, in all three formats.
• feat: sum_forward, the window sum_back already was pointed the other way #56 takes the sum_forward half — the builtin, sum_forward_partitioned.yaml, the generator row, and the comment reword to "the two windows" — so that operator arrives complete instead of landing and then getting a by= bolt-on.

Nothing from this branch is dropped.

---

Generated by Claude Code