Return an empty array from chained vector lookups that select no rows

[MaxGhenis]

What this fixes

A chained vector parameter lookup that selects no rows raises once it reaches numeric leaves. The audit's witness is a tree with one leaf a.b.c = 0 and three empty key arrays:

node = ParameterNode(data={"a": {"b": {"c": {"2020-01-01": 0}}}})("2020-01-01")
none = np.asarray([], dtype=str)
node[none][none][none]
# IndexError: index 0 is out of bounds for size 0  (vectorial_parameter_node_at_instant.py:416)

After the empty selection every numeric child holds zero values. The "scalar per child" path, meant for 0-d and length-1 children, read the first value of each. That path now takes only 0-d and length-1 children. Length-0 children go to the existing N-element path, which stacks the children into a (K+1, N) matrix and returns a length-0 float array.

The independent review found two more cases in the same code, now fixed:

• Unknown keys after an empty selection. A one-element key broadcasts over zero rows, so node[none][["missing"]] returned an empty array. The unknown-key check only fired when the result held NaN, so nothing raised. It now also fires when any key index is the missing-child sentinel, and the lookup raises ParameterNotFoundError as it does for rows that exist.
• Nodes (structured levels) after an empty selection. Zero rows took the one-row path. A valid one-element key returned one row of NaN, and node[none][["b"]][["c"]] read [nan] (master too). Zero rows now give zero rows at nodes as at leaves, broadcasting the key the same way: no rows for a one-element key, and a shape-mismatch error for a key of two or more.

Invariants

For any parameter tree of depth 1 to 3 (1 to 3 children per node, numeric leaves) and any 0 to 20 paths through it:

• Looking the paths up one level at a time with key arrays gives an array of shape (number of paths,). Each element equals the scalar lookup of its path. Zero paths give an empty array.
• A key that names no child raises ParameterNotFoundError at any level, among valid keys for the selected rows, or alone when no row is selected.

Tests

• test_fancy_indexing.py::test_empty_fancy_indexing_at_every_level (leaf, node and triple lookups on the existing fixture tree) and ::test_empty_fancy_indexing_on_a_single_path (the audit witness).
• test_fancy_indexing.py::test_one_key_after_an_empty_selection: one-element keys after an empty selection, at a node and at a leaf.
• New test_fancy_indexing_property.py: the two properties above, 300 examples each.

Before the fix, on master's policyengine_core/parameters:

test_fancy_indexing.py:            2 failed, 8 passed   (IndexError: index 0 is out of bounds for size 0)
test_fancy_indexing_property.py:   1 failed             (minimal: depth 2, one leaf, no paths)

Round 2 (after the independent review), the new tests on the first head f8426179:

test_fancy_indexing.py:            1 failed, 10 passed   (node[none][["b"]] had 1 row: assert 1 == 0)
test_fancy_indexing_property.py:   1 failed, 1 passed    (DID NOT RAISE ParameterNotFoundError)

After:

test_fancy_indexing.py:                   11 passed
test_fancy_indexing_property.py:           2 passed
test_parameters.py:                       14 passed
test_numpy2_structured_arrays.py:          4 passed
test_vectorial_parameter_node_copy.py:    64 passed

Ruff format and ruff check pass. I ran these files one at a time and left the full suite to CI, because the shared host is saturated.

Impact

No country output can change. The only inputs that take a different path are selections of zero rows, which used to raise or return NaN rows, and unknown keys, which now raise in one more case. Children with one value, which includes every first-level lookup, still take the scalar path. Children with two or more values still take the N-element path. A microsimulation never looks up parameters for zero rows, so no A/B run is needed.

axiom: n/a: engine fix in vector parameter lookup, no policy change

Other audit findings, handled elsewhere

The 2026-10-06 invariants audit of master 757147c7 also confirmed defects that this PR leaves alone, because other work owns them:

• Stock ADD changing later plain reads: draft Cache ADD and DIVIDE results only where a plain read returns them #571.
• Year deletion leaving monthly fast-cache entries: draft Drop every contained period from the fast cache when deleting or setting a period #566.
• Disk deletion of contained periods, and disk branch-name parsing and prefix collisions: draft Delete the periods within a deleted period from disk storage #565 (it inherits Keep branch period reads, dumps, and disk deletion consistent #552).
• Clone aliases bound to the parent: fixed by Bind a clone's calc and df to the clone, and give it its own system's entities #586.
• Serialization and subsample input provenance: Restore inputs as inputs, evict the fast cache on holder writes, reject non-numeric uprating and defined_for #576 and Rebuild a subsampled simulation from its inputs, not its calculated values #573.
• Restored entity counts: Fix restore_simulation entity counts #548.
• A monthly input leaving a cached annual sum stale (set_input for one period leaves the variable's cached sum or twelfth over an overlapping period in place #579): tracked separately.

Branch staleness after set_input and raw-file restore are documented behaviour, not defects.

🤖 Generated with Claude Code