Derive mapped-node labeling at compile time

[SimonHeybrock]

Spike supporting the discussion in scipp/sciline#237 (demand-driven generic providers) — see the design document there for context. Draft for discussion.

What changes

Graph.map no longer relabels the descendants of mapped roots as MappedNode. The stored graph keeps original node names at all times; which nodes carry which indices is derived inside to_networkx() from reachability of the mapped roots plus recorded reduce specs, and the MappedNode representation now exists only transiently during compilation. The observable output of to_networkx() is unchanged.

The practical consequence: map() commutes with adding branches to the graph. Sciline needs this so that generic providers can be instantiated on demand after mapping — the mapped-labeling of late-added dependents falls out of the derivation instead of requiring eager instantiation before map().

reduce() records what it consumes (index/axis/all, plus groupby's extra index) in a side table; the remaining indices of the reduce node are derived at compile time. Validation still happens eagerly at call time against the currently derivable indices.

The shadowing seam

Idioms like reduce(name=C, index='x') and pipeline[C] = pipeline[C].map(...).reduce(func=merge) give the reduce result the same public name as the mapped node it reduces. Previously this worked because the mapped node was labeled MappedNode(C) while the result was plain C — the dual identity was implicit and global. With plain names it is a genuine collision. The dual identity turns out to be irreducible exactly at these seams, so it is kept there and only there: when a reduce result or an assigned branch shadows a mapped node, the mapped node is relabeled to an explicit MappedNode alias. New accessors (node_indices, named_indices, value_keys) expose derived indices to callers that previously scanned for MappedNode labels (sciline's get_mapped_node_names).

_from_orig_key and the eager relabeling helpers are gone; __getitem__/__delitem__/__setitem__ lose their MappedNode special cases.

Found along the way

graph1['c'] = graph2['d'] on main merges the branch's own 'c' node with the destination during sink-renaming, producing a self-loop (reproducible on main; invisible because nothing topologically sorts the stored graph). The index derivation here tolerates self-loops to stay behavior-compatible; the underlying __setitem__ issue deserves a separate fix.

Verification

All 143 tests pass unchanged. mypy: one error fewer than main (22 vs 23 pre-existing). The sciline side of the spike (all 245 sciline tests green, forward-instantiation hook removed) is in scipp/sciline — branch 235-deferred-mapped-labels.

🤖 Generated with Claude Code

[SimonHeybrock]

Follow-up per discussion (a046ffb): the MappedNode alias mechanism is removed in favor of strict name uniqueness — reduce() and __setitem__ now raise clear errors when the new node would shadow a like-named node inside the branch (the pipeline[C] = pipeline[C].map(...).reduce(...) idiom and reduce(name=<reduced node>)). This also turns the silent self-loop on main into an error, and deletes named_indices plus the self-loop tolerance (net −47 lines vs. the previous state of this PR). Grafting a mapped branch at a new node name is now allowed, since indices follow from derivation. The explicit escape hatch for subgraph reuse would be a future Graph.rename.