Prototype: demand-driven generics without map-time forward instantiation

[SimonHeybrock]

Q1 spike from the design document (docs/developer/architecture-and-design/demand-driven-generics.md): remove the map-time forward-instantiation hook by relying on cyclebane deriving mapped-node labeling at task-graph build time. Depends on scipp/cyclebane#32 — CI will be red until that lands and is released. Based on the #237 branch; draft for discussion.

What this shows

With cyclebane keeping plain node names and deriving which nodes carry which indices at to_networkx() time, map() commutes with provider insertion: generic providers instantiated on demand after mapping receive their indices from the derivation. Consequences here:

• The map() hook and the entire seed-restriction machinery are gone; forward chaining survives only behind output_keys(), whose fate is Q3 (rule-graph inspection).
• Mapped roots count as satisfied keys, so backward instantiation never tries to provide them.
• reduce() with an explicit key treats the key as a demand and instantiates providers for it. Without a key, the reduced sink is resolved among nodes actually present, so reducing a rule-derived sink requires an explicit key — two tests updated accordingly. This is the user-visible rule to review.
• get_mapped_node_names reads derived indices via cyclebane's new named_indices accessor instead of scanning for MappedNode labels, and loses the multiple-candidates ambiguity handling (plain names are unique).

All 245 tests pass (against the cyclebane branch); notably, before this change, the whole suite already passed against the cyclebane branch without any sciline modification, i.e. the cyclebane change is transparent to current sciline.

Discussion points

• The reduce-key rule above: acceptable, or should sciline defer sink resolution further?
• The shadowing seams (reduce results and grafted branches reusing a mapped node's name) keep a scoped MappedNode alias inside cyclebane — see the discussion in Derive mapped-node labeling at compile time cyclebane#32.

🤖 Generated with Claude Code

[SimonHeybrock]

Follow-up per discussion (4cbf9fd): forward chaining is now fully deleted (_instantiate_forward, forward_bindings) — the engine is backward-only. output_keys() returns concrete sinks plus return-type patterns of generic providers not consumed by other generic providers; no-arg visualize() shows the concrete part of the graph. With scipp/cyclebane#32 forbidding the shadowing idioms, get_mapped_node_names loses the multiple-candidates disambiguation (index_names= is pure validation now) and the affected tests use distinct names for a mapped node and its reduction. Decisions recorded in the design document.

[SimonHeybrock]

Merged into #237 (fast-forward, commits preserved), which is now the consolidated end-state proposal against main. See the design document there for the decision log.