Floor time: dissolve the #6848 once-per-entry fixpoint + reconcile-miss qualified-fill (~18min residual)

[briansrls]

Summary

Follow-on to #6999 (bisect of the #6848 namespace-walk CI-floor regression). #6999 landed an
entry-closure memo + reconcile call-order fix, but batch-4 wall time stayed pinned at
~1111–1125s (unchanged from pre-#6999) — the ~18min residual named in dashboard work item
adhoc-21c65e1a-2ff.

Root cause: SelfHostReadsRealBytesGate, SelfHostStalenessGate, and EmitHostGate
declared heavy_whole_tree_resolve: false in gate_runnable_profile
(src/v2/workflow/ci_floor_plan.dag). claim_executor::run_walk routes false-flagged batch
units onto freshly thread::spawn'd threads → resolve_entry_graph → process_shared_index, a
thread_local! MultiEntryIndex keyed by (thread, roots). Each spawned thread gets a cold
index, so #6999's entry-closure memo and pool_qualified_fill cache — both living on that
per-instance index — are built and discarded on every batch-4 gate run. DagCompileCleanGate,
SourceRootIngestGate, and RegenVerifyGate already declared heavy_whole_tree_resolve: true
and run on the shared main-thread walk_memo path instead — these three gates were simply an
inconsistent application of an already-established pattern.

Fix: flip heavy_whole_tree_resolve: false → true for the three gate profiles (3 lines,
.dag metadata only — no Rust changes, no resolver/binding semantics touched).
floor_heavy_resolve_chain_resource_edges already serializes all heavy_whole_tree_resolve
gates pairwise (witness_plan_serializes_heavy_resolves), so the three newly-heavy gates land
in their own batches sequentially on the main thread — sharing the process-wide
walk_memo/process_shared_index cache instead of paying a cold walk each, without co-residing
in one batch.

Diagnosis + mechanism write-up: docs/plans/floor-time-namespace-walk-regression-diagnosis.md
§6.

Out of scope: resolver semantics, extend_sources_to_both_closure_fixpoint/
pool_qualified_fill build logic, rc_map_insert quadratic, UnlistedImportUse suppression,
and batch-2's separate ~301s discovery-side gap (a different scheduling shape — stays open on
the work item).

Test plan

• cargo build --release --bin claim_batch --bin claim_executor: green (local, sccache bypassed
  due to a container resource-contention issue unrelated to this change).
• claim_batch --source-root dag --source-root src/v2 --entry src/v2/test/claim/ci_floor_plan_witness_test.dag --functions <all 21 zero-arg witness/falsifier fns> --hermetic: 21/21 PASS, including witness_plan_serializes_heavy_resolves,
  witness_optin_plan_serializes_heavy_resolves, witness_optin_emit_host_isolated_from_corpus,
  witness_emit_host_profile_forbids_corpus_co_residence.
• Grep-confirmed no existing test/lens hardcodes heavy_whole_tree_resolve: false for these
  three gates; runnable_excludes_corpus_co_residence gates on profile.memory, orthogonal to
  this flag.
• Pending (by-execution, not yet observed): a real CI floor run showing batch-4 wall time
  move off the ~1111–1125s baseline toward the ~342s pre-namespace wave 1: containment-tree resolution — layered census + name-derived loader (salvage) #6848 class, per governor-mandated
  measurement protocol (gh api job logs, claim_executor: batch N timestamps).
• Pending (governor-added acceptance criterion, memory-beside-time): compare the governor
  receipt line (peak_current, hard_backoffs, forced_serial, budget_exceeded) on the same
  CI run against a same-host-class baseline. Capped hosts already run 14–16GB peak against a
  16.1GB budget; routing 3 more gates onto the persistent shared main-thread index changes the
  residency profile even though total resolve churn should drop. A material move toward the
  budget is a stop-and-report, not a ship.

Closes dashboard work item adhoc-21c65e1a-2ff.

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.

[gunbai-bot]

CI failure at 58e1c18 (batch-2 discovery: doc_graph_has_no_orphan_docs) was unrelated to this PR's batch-4 heavy_whole_tree_resolve flip — the origin/main merge (#7023) brought in docs/probes/emitter_residual_site_map_2026-07-21.md with no inbound link. Fixed by adding a bind: provenance line at its topical home (the ^emit_representation_mismatch milestone it documents, dag/gunbc/v1_deletion_plan.dag), rather than skipping or widening the orphan check. Verified locally: doc_graph_has_no_orphan_docs / _has_no_dangling_links / _universe_is_nonempty all PASS, and the batch-4 heavy-resolve witnesses (witness_plan_serializes_heavy_resolves, witness_optin_emit_host_isolated_from_corpus, falsifier_plan_structure_holds, etc.) all still PASS.

— sent from proud-bear-438