Native-realize the dependency-fact composition: census 180,600ms failing to 68,224ms passing

[briansrls]

The partition-closure census exceeded its long-lane ceiling (180,600 ms against 180,000 ms). This native-realizes the composition that was the cost.

Result

	before	after
dependency_resolution_facts_live	112,261 ms	216 ms
w_live_corpus_partition_closure_holds	180,600 ms — failing	68,224 ms — passing
Long-lane ceiling	180,000 ms	unchanged

62% under the ceiling. No budget raised, no cadence moved, no witness relocated. The census landed within 0.5% of what the localization predicted (68,555 vs 68,224).

The cause, measured rather than reasoned

Per term, one process, live corpus:

import_resolution_facts_live                      146 ms
reference_derived_import_resolution_facts_live      5 ms
union + dedup of the two (interpreted)        104,943 ms
dependency_resolution_facts_live (all)        112,261 ms

The host leaves are 151 ms of 112,261 ms. The interpreted composition was the entire term, and effectively the entire census. So this moves the composition into the host rather than caching the leaves — caching them recovers 151 ms and leaves the census over its envelope.

Four hypotheses refuted by execution first

Recorded so nobody re-runs them:

• Duplicate (importer, target) work in the consuming index — raw == unique exactly, measured post-join at the module grain, with path→module injectivity measured rather than assumed and zero silently dropped edges. The facts are already deduped at the same triple, so a dedup there had no inputs to act on.
• The interpreter Map primitive — map_insert 38 ms at n=2000, 68 ms at n=8000 against a take-only subtractive control: 1.8× for 4× the items. Not quadratic. Explicitly not claimed as a defect; one population point cannot establish asymptotics. Left as quarry.
• The bucket's list_snoc_item — replaced with a cons, measured 105,007 ms against a 104,943 ms baseline. No change. Reverted.
• The read-hoist recorded in stage0_partition_closure (173.6 → 177.6 ms, noise) — its stated explanation is corrected below.

Preserved, and why each matters

• Exact identity on {path, import_module, target_declared} — an unresolved target is a target_declared: false row, and the census exists to judge those.
• Reference-first order, first occurrence — order is observable: module_target_index_from conses targets in arrival order, and module_impact_query derives provenance from which copy of a dual-backed fact survives.
• The two-producer seam — the leaves remain the swappable pair dependency_edge_source_migration_note designates for the namespace terminal step. The native arm composes them; it does not replace them with a black box.

Evidence

• Live-corpus ordered parity, row for row, against the interpreted spelling, plus an independent target_declared distribution check — banked at 7827977bd4 before that spelling was deleted, so "faster" provably is not "different".
• 5 synthetic Rust controls on the dedup law (cargo test union_dedup_import_facts_law, 5 passed). The load-bearing one is declaredness_variants_of_one_pair_are_distinct_facts: a dedup keyed on (path, module) alone passes every other assertion and silently destroys the census's subject.
• The enrolled control already existed — module_graph_edge_source_witness_test points the production function at a four-module fixture and asserts the union at identity grain in both directions, so it now exercises the native arm and reds on a dropped reference-only edge or a surplus.

Deleted rather than kept dormant

union_import_resolution_fact_lists, dedup_import_resolution_facts, add_unseen_import_fact_indexed, import_facts_bucket, ImportFactDedupAcc, import_resolution_fact_eq. A dead 105-second implementation beside the live one is future ambiguity, not authority. dependency_edges_from_import_facts and import_fact_to_dependency_edge are kept — still consumed by reference_derived_resolution_facts_flagged_pending_reference_producer, the latter also named by module_impact_query as the future provenance-stamp site.

Notes corrected, because they were actively misleading

stage0_partition_closure explained its null hoist result by asserting "the live fact producers dominate and are already memoized host-side". Never measured, and false — that sentence pointed the next reader at the one part of the pipeline that was already cheap. The live-census row's own 1.3%-margin warning came true at 180,600 ms and is now closed with the measurement. dedup_import_resolution_facts_cost_note keeps its name deliberately: stage0_partition_closure and repo_atlas_projection cite it by symbol, and deleting it would rot both citations. module_impact_provenance_derivation_note is repointed — it depends on the reference-first law, which changed home, not content.

Explicitly out of scope

The selection-graph twin is untouched. Sharing the helper is correct under §3, but selection_edges feeds selection_adjacency, which decides which witnesses run; reordering an input without proving order-insensitivity risks the empty-observation-narrow class on the affected-set path. That consolidation needs its own equivalence PR. Also excluded: floor-snapshot reuse, cache-key design, Map-primitive investigation, budget changes, other falsifier failures, G2.

Bootstrap ordering, recorded because nothing in the tree states it

Regen runs the existing binary, so a .dag already calling a new builtin fails regen with function not found in scope; but the new binary cannot compile until the generated dispatch carries the arm, since the macro rejects an unrostered arm. The order is forced: Rust + roster → regen → build → .dag adoption.