[trissim]

R1e · guards 2/2 · production −1410 +241 · tests −4 +285

Owner requirement: every one-time cost is paid in function-catalog preparation, before the server reports ready; nothing is JIT-compiled afterwards in the server, its workers or the client. Receipt with method, tables and certificates: docs/refactor/consolidation-20261009/R1e-no-jit-after-readiness.md.

Measured on main (cold Numba and catalog caches, CPU-only, Numba's numba:compile event and cache lookups counted per process)

• ExampleFly and ExampleColocalization compiled 7 kernels inside timed execution (0.91 s, 0.77 s): the Costes warm-up passed only anti-correlated data, and linear_costes branches on the data.
• 52 of 298 first-party kernels had no compiled signature anywhere after a full preparation.
• The client loaded 60 to 178 kernels (0.5 to 1.0 s) converting a .cppipe, because building a callable's contract ran its preparation hooks. It never runs a kernel.
• The server first-imported 106 modules across the first request, and each worker repeated them (0.09 to 0.13 s; 0.44 s when PolyStore's registry cache missed).
• Not reproduced: R1c's "workers compile thresholding kernels". Workers are forked before readiness with every prepared kernel loaded.

Changed

• Every such kernel is now reached by its owner's existing preparation (branch, dimensionality or dtype the warm-up missed), or deleted because nothing calls it (30 kernels). No list of kernel names; no kernel body changed.
• CallableProjection.declared_contract() reads a contract without preparing; .cppipe conversion and pattern lowering use it. CompilationSession prepares through normalize_prepared_function_pattern, so a process without catalog preparation still prepares at compile time. Signatures are identical before and after preparation for all 1035 catalogued targets.
• The catalog server builds the storage backend registry and completes the domain-extended registries before the worker fork.

After

	main	branch
Compilations after readiness (Fly / Colocalization)	7 / 7	0 / 0
Kernel loads in the client	60 to 178	0
Kernels with no compiled signature after preparation	52	0
Server imports after readiness, setup / compile / execute	44 / 2 / 60	0 / 2 / 57
Pipeline compile first / second, ExampleHuman	0.302 / 0.282 s	0.181 / 0.176 s
Cold preparation, traced, mean of six	52.3 s	51.0 s

Wall times were taken with other jobs running on the machine; counts are exact.

Guards

tests/integration/test_no_jit_after_readiness.py: one cold run of two pipelines over a real server in a child process. (1) No process compiles or loads a kernel after its server's readiness; a client never does. (2) Every openhcs Numba dispatcher that is not inline="always" has a compiled signature at readiness; seen failing with one warm-up removed. Integration tier (a cold preparation takes about 90 s). They cannot see a kernel prepared with one signature and reached with another.

Gates

• verify_branch.sh r1e: 11656 passed, 34 failed, 2 errors; new against the known list: 0.
• Official30 parity under heavy_slot.sh and official30.lock, cold: 1 passed in 716 s (all 30 cases); its trace shows 0 compilations after the first preparation in any process.
• smoke_local.sh rc=0; guardrails_local.sh rc=0; guards 2 passed.

For the reviewer

• Three Numba overrides were removed for inputs production does not pass (block_labels, non-boolean fill_labeled_holes_below_size, 3D relabel_sequential); those inputs now take the inherited NumPy/SciPy implementation. Certificate: tests-only. The alternative is to keep them and add warm-ups.
• intensity_zernike_moments_batch (an exported name with no caller) is deleted.
• Remaining late work, 0.03 s per process: PIL plugin and imagecodecs loading at first image I/O. Its owner is PolyStore; Storage backends prepare their process-wide lazy state on request PolyStore#40 adds the preparation, to be adopted after a release.

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.