Colony & Microfluidic Phenotype Quantification in_progress

Does whole-cell E. coli phenotype survive colony embedding, and match microfluidic-device data, across a cost/fidelity ladder of cell models? Runs N whole-cell (or cheaper) agents in one pymunk 2D device geometry and measures a shared phenotype panel (growth rate, size-at-division, added length, inter-division time). Part A (colonies-01..03) is the enabling compute foundation — correct division, characterised cost, bounded memory on Linux; Part B (colonies-04..09) is the phenotype science.

Question. Do the emergent single-cell phenotypes of the whole-cell E. coli model — adder size-homeostasis, the inter-division-time distribution, and the growth rate — survive embedding into a physically-packed, dividing microcolony, and do they match microfluidic-device data, across a cost/fidelity ladder of cell models (simple viva-munk agent -> growth surrogate -> full 55-process WCM) run in the canonical devices (mother machine, daughter machine, free colony)?

Hypothesis. Single-cell phenotype is a property of the cell model, not the container: the WCM's adder size-homeostasis, inter-division-time CV, and growth rate are preserved when cells are embedded in a pymunk 2D colony that grows, packs, and divides, and they are reproduced (to tier-appropriate fidelity) by the cheaper simple-agent and growth-surrogate tiers. The compute foundation needed to run this — daughter hydration through the engine, natural division, coherent physics, and bounded memory — is infrastructure, not the result.

The object is a growing microcolony of E. coli, every agent embedding the full whole-cell model (the 55-process baseline that grows, transcribes, translates, replicates its chromosome, and divides) in a 2D pymunk environment that resolves physical packing. A cell elongates as its dry mass accumulates to the division threshold (~one cell cycle), then splits into two daughters placed along its long axis that reset and regrow (see the colony-animation GIF). No biological model change was made; the compute work is infrastructural — daughter hydration, division counts and placement, realistic physics, streaming a bounded per-cell phenotype panel to zarr, and characterising the compute and memory cost of running many whole cells at once. The scientific payload is the phenotype quantification: does the whole-cell model reproduce the size-homeostasis (adder), inter-division-time, and growth-rate statistics that microfluidic devices measure — and do cheaper cell-model tiers reproduce them well enough to substitute at colony scale?

Main claim. Whole-cell single-cell phenotype (adder size-homeostasis, inter-division-time distribution, growth rate) is preserved under colony embedding and division and is quantifiable across a model-tier ladder and device geometries. The compute foundation that enables this is single-machine-correct, cost-flat, and memory-bounded on the Linux HPC target — the memory growth earlier read as an HPC blocker is largely a macOS allocator artifact.

Caveats

• Single-machine, single-seed (0) compute measurements; cross-node scaling is a follow-up.
• Part B phenotype distributions are held for comparison against experimental microfluidic data; the surrogate tier and real mother-machine data are planned.
• The ~700 cells/node budget assumes ~11 cells packed per Ray actor (sim_data shared within the actor). One-cell-per-actor (max GIL parallelism) makes each actor pay a full ~888 MB sim_data baseline and is RAM-bound at ~215 cells/node. The right packing is a deployment decision, not a fixed number.