feat(asap-aware-mapping): rebuild applicability.rs as a view over search's PlanSpace (#257)

[zzylol]

Closes #257. Part of #33. Stacked on #263 (issue #252) and #259 (issue #251) — do not merge before those two.

What changed vs. the old rule-based traversal

crates/asap-aware-mapping/src/applicability.rs (PR #247, unmerged) re-derived "is optimization X applicable" as its own yes/no fact, walking the tree a second time per rule (SketchApplicabilityRule wrapping boundary::implementation_for_with, SharedSubexpressionRule wrapping share_common_subtrees). Now that search::search_workload (#252, over the ReplacementStrategy extension point from #251) already computes, per site, every candidate replacement any registered strategy proposes, this module is rebuilt as a pure view over that PlanSpace:

• The reframing: "applicable" now means "this site's candidate list contains something other than the trivial/no-op realization."
  • OptimizationKind::SketchApproximation — the site has at least one Replacement::Summary candidate that actually realizes a sketch family (not just the exact/pass-through candidate the boundary would commit to anyway).
  • OptimizationKind::SharedSubexpressionReuse — the site has consumer_count >= 2 and its candidate list contains SharedSubtreeStrategy's "build once and share" candidate.
  • Each finding's reason is literally the matching candidate's own rationale string — no new prose invented to re-explain why a candidate is valid.
• ApplicabilityFinding/OptimizationKind/find_applicable_optimizations/find_applicable_optimizations_with keep their exact public shape and contract (same fields, same two variants, same #[non_exhaustive] discipline). find_applicable_optimizations_with now takes a &[Box<dyn ReplacementStrategy>] (mirroring search_workload_with) instead of a rule list — the natural consequence of strategies replacing rules as the underlying data source.
• Deleted, not kept alongside the new implementation: SketchApplicabilityRule, SharedSubexpressionRule, collect_sketch_findings, register_site, walk_rc_children, for_each_operator_child, and the ApplicabilityRule trait itself. ApplicabilityRule is redundant with ReplacementStrategy once findings are a view over PlanSpace — a new optimization needs a new impl ReplacementStrategy wired into search::default_strategies() regardless (that's the only way its candidates reach the PlanSpace this module reads), so a second, applicability-specific extension point would just duplicate that one. Documented in the module doc, mirroring replacement.rs's own "why this exists alongside boundary and bind" style.
• One small new traversal kept: collect_locations, whose only job is turning "this Rc" into human-readable breadcrumb text (root "dash_a" > lhs) for ApplicabilityFinding::location — PlanSpace/MemoGroup deliberately track only Rc pointer identity, not caller-facing prose. It makes no applicability decision of its own; it runs the same regardless of what any strategy found.
• Updated lib.rs's module-level doc (added the applicability bullet, and a search bullet that had gone missing from an earlier PR) and added a new "Applicability reporting" section to docs/asap_aware_mapping.md. Fixed a handful of now-stale cross-references in replacement.rs/search.rs that pointed at the deleted PR feat(asap-aware-mapping): optimization-applicability rule framework #247 traversal.

Old guarantees re-verified against the new implementation

The two properties the old traversal's docs specifically called out are both still upheld, by mechanisms now living in search.rs rather than here:

1. Maximal finding, not one per subsumed descendant. search::discover_sites only recurses into a node's children the first time that node's Rc is seen (exactly register_site's old discipline) — a node nested under an already-discovered shared ancestor only becomes its own site if something outside that ancestor also references it.
2. A shared node is one finding, not one per path. MemoGroups are keyed by Rc pointer identity, so there is exactly one group (hence at most one finding) per distinct Rc, regardless of how many paths reach it.

a_shared_sketchable_aggregate_is_reported_only_once (pinning guarantee 2) passes unchanged. Added a_deeply_shared_subtree_under_different_parents_is_reported_once as a further regression test for guarantee 1 (mirrors search::tests::nested_shared_subtree_below_an_unshared_parent_is_still_discovered).

All other existing tests (positive/negative fixtures for both sketch and reuse findings, custom-strategy plumbing) were preserved with only mechanical updates (e.g. SketchApplicabilityRule::new → SketchFamilyStrategy::new passed as a strategy).

Checks

cargo build --workspace --all-targets, cargo test --workspace, cargo fmt --all -- --check, and cargo clippy --workspace --all-targets --all-features -- -D warnings all pass clean.

🤖 Generated with Claude Code

[zzylol]

Superseded — its content (applicability.rs as a view over PlanSpace) was folded directly into #259 (feat/replacement-strategy-251) rather than staying a separate stacked PR.