[Xubqpanda]

Importing a configured 134-task ALFWorld pool eagerly built every SETUP, and each called resolve_execution: full venv hashing plus a Python probe. Make ALFWorld, Jericho, ScienceWorld and DS-1000 declarations lazy so key enumeration does not construct setup/verifier values.

Registry declaration batches verify each execution artifact once and reuse metadata copies within that operation. Provider snapshots retain the artifact admitted at entry instead of re-admitting it per task. Actual provider RPC startup, actor/service/verifier launch, input consumption and score-reuse admission still validate freshly. No process-global permission cache is introduced.

The repository scan also fixes full-verifier traversal for inventory kind reporting, repeated native-research manifest/index construction, repeated SpreadsheetBench native inventories during planning, shared source/JAR/ROM reads, frozen-session catalog subprocess repetition, and a duplicate TravelPlanner world check. World/database byte checks remain fresh per selected task; only the immediate duplicate is removed. Source hashing is scoped rather than relying on process-wide timestamp caches, including filesystems with coarse timestamps.

Controlled evidence: the same authored 134-task pool produces 134 runtime resolutions at import before the change and 0 after it. A 134-lookup metadata batch admits once; a subsequent scope and explicit launch admission still recheck, including a damaged-artifact control. No real ALFWorld GB-read or wall-time improvement is claimed.

Validation:

• Dataset/runtime/provider regression: 238 passed, 43 environment-dependent skips.
• CI guards, wheel resources and clean-install checks: 193 passed.
• Final snapshot API/entry-admission regression: 50 passed.
• Focused performance controls: 13 passed, covering lazy pool imports, selected-only construction, scoped admission, catalog copies/config changes, manifest/native inventory reuse, restored-mtime source edits and fresh world checks.
• Scan covered 1,522 tracked Python files plus Shell/JS/TS checks. No released benchmark or model run, and no shared service changes.