perf(r2rml): F20 RefObjectMap-target resolution prune — q031 72s→5.2s (kill the shared-predicate fan-out)

[aaj3f]

F20 — RefObjectMap-target resolution prune (kill the shared-predicate fan-out)

Stacked on perf/r2rml-pr4d (#1501). Kills q031's over-scan: a star subject ?p bound solely as one RefObjectMap object (parent class C) has its downstream single-predicate star's TriplesMap resolution constrained to C, so ?p edw:name ?pn no longer fans out to all six edw:name-bearing dims.

q031 — honest numbers (cache-thrashed, full-corpus 3-rep protocol)

	load_table.n	scan_table.n	rows	wall
OFF (pre-F20 fan-out)	7	1448	5000	154.0 s
ON (F20 prune)	2	243	5000	5.21 s (3-rep median, in-protocol)

72 s → 5.21 s (~14×), a deterministic fan-out kill (7→2 loads, 1448→243 scans) — the load/scan collapse is cache-independent; the wall is the measured cache-thrashed 3-rep median. rows_only parity (row count ON==OFF; hashes differ by unordered-LIMIT nondeterminism, expected).

Residual (measured, not inferred — doc 20 §8)

The 243 surviving scans decompose (per-scan trace, reverted) as 2 fact + 1 deduped DIM_PRODUCT parent lookup + 241 budgeted DIM_PRODUCT MAIN re-scans of ?p edw:name (row_budget=Some(5000) alone disqualifies the scan_cache; filters=0, not parent lookups). An A/B differential collapsing those 241→3 measured Δ −0.5 s — they cost nothing. A query-scoped small-dim main-scan cache is therefore measured-and-rejected. q031's residual is the fact-scan + materialize + loadTable floor — the same floor as q016-post-#1501; the isolated first-ask still pays ~2–3 s of loadTable (amortized to load=0 in corpus order). Named levers to close ≤3 s (not in this PR): the loadTable-METADATA cache + PR-2a / materialization.

The verdict chain (six evidence-forced flips — preserved in the register)

q031 cost six flips: PR-5-suspected → bisect-refuted → memo×limit-pushdown → cold-floor loadTable-pin → measurement: RefObjectMap-target fan-out (F20) → the residual measurement above (small-dim cache rejected). Along the way F19 (the "query-scoped parent-memo doesn't engage" theory) was refuted — the memo already shares on q031's path (with_active_graph clones the Arc; confirm-experiment 1× shared vs 5× fresh-ctx control); its one sound residual fix (with_graph_ref memo-clone) is filed and deferred, affecting no corpus query.

Soundness (the load-bearing part)

The prune reuses the existing class_prune_hint verbatim (zero operator change). It fires only when BOTH:

• Invariant A (join-var provenance): ?p's only binding in scope is exactly one RefObjectMap member (unique parent class C). Computed via Pattern::produced_vars (not referenced_vars) — so a BIND target or a path endpoint (which rebind ?p) decline, while FILTER/MINUS/(NOT) EXISTS (which only constrain) do not. One principled choice; future IR pattern types inherit correct treatment. Declines on: a second object binding, a UNION/VALUES-bound ?p, a second RefObjectMap with a different parent, a mixed ref/non-ref predicate, an ambiguous parent class, or a standalone class assertion on ?p.
• Invariant B (template-disjointness): wildcard_class_fusion_is_safe(C) — the F10 vertical-partition guard, verbatim; a map sharing C's subject template that also maps the predicate declines.

Cross-scope pre-bound ?p (seeded outside the scope) is sound by conjunction: the in-scope ref member restricts ?p to class-C FK objects regardless of the prune (engine-tested via CountingProvider: valid product 1→1 row; non-product Store IRI → 0 rows, identically pruned or not).

Kill switch: FLUREE_R2RML_REF_TARGET_PRUNE (default on) — off is byte-identical to the pre-F20 fan-out, as is every DECLINE path.

Tests + gate

12 tests: 10 hermetic rewrite (positive + BIND/UNION/VALUES/2nd-binder/different-parent/template-share/class-assert declines + FILTER-no-decline) + 2 engine-level (cross-scope conjunction). Full fluree-db-query lib green (1247). Full-corpus baseline at this head: 54 records, 0 hash mismatches, 0 perf violations; the 3 expected-error queries unaffected; 44/51 ok-queries ≤3 s.

CI note: this PR's base is perf/r2rml-pr4d, not main, so repo CI does not fire on it (stacked-PR convention).

[aaj3f]

Consolidating at maintainer request. #1502 (F20 — the RefObjectMap-target resolution prune) is carried forward in #1507 (the R2RML/Iceberg virtual-dataset performance program), which brings the #1475–#1507 lineage forward as a single reviewable unit. Its verification of record is the program's corpus benchmark, published as C2-bench-wave1.md (native == virtual, 0 hash mismatches). The pre-consolidation branch tip is preserved at tag archive/pre-refactor-2026-07-21/perf_r2rml-q031-refprune — no history was rewritten. Closing in favor of #1507; discussion continues there.