[trissim]

A fresh client-owned execution server currently builds the complete function catalog before binding, then waits for it again before running a pipeline. Converted CellProfiler pipelines import their functions directly, so that catalog work adds startup latency without helping their execution.

Bind the endpoint first and prepare the catalog on the first catalog request. Before evaluating pipeline source, inspect its imports and initialize the registry when it needs a registry-created virtual module; direct imports proceed without that global projection. Concurrent catalog and pipeline requests share one metadata load. This keeps catalog reads and external virtual imports available while shortening fresh execution startup.

On the same CPU-only integration checkout, the official30 one-well fresh-server sweep passed 30/30 and all 110 produced result files matched the control byte for byte. The sum of per-observation totals fell from 311.858 to 274.266 seconds (−37.592; all 30 cases faster, median −1.171 seconds). Non-job time fell by 58.822 seconds; compilation rose by 25.126 seconds as some registry work moved there. The execution-job difference is run variation and is not claimed as an effect of this patch. A live client fetched all 266 catalog functions on demand after connecting.

Raw CSVs, native-relative figures, and timing boundaries are included. Validation: 104 focused tests passed on this PR branch; Black and git diff --check passed. Ruff is not installed in the local benchmark venv.

Refs #165. This PR is stacked on #167 because it reuses the registry inventory lock added there. The benchmark checkout also contained other pending performance changes, including the direct-docstring dependency update in #175; only the lazy-catalog code changed between the paired CSVs.

[trissim]

Issue #179 is being fixed by the blind-analysis integration owner on fix/compile-observation-identity-20260929. zmq_execution_server.py changes are confined to compilation_signature on ZMQExecutionContext, _resolve_compilation, and _store_compile_artifact. No function-catalog startup/preparation edits; your lazy-catalog ownership stays separate. _export_runtime_observation is unchanged (PR #157 owns that path).