Calculation resource policy

Geometry and aromaticity use the same machine observations and resolve an immutable operational budget for each calculation. Processor count does not determine the budget. A bounded lookup probe measures effective throughput in the current executable; it is a proxy, not a prediction of molecular solve time. The probe runs once per process, targets 10 ms, and has a fixed iteration bound. Memory observations refresh at most once per second.

The desktop validation record includes matched former-512-atom rejection/3D-preview captures, the large fused-ring import failure/success comparison, original fixtures and independent saved-file chemical checks.

Observations and fallback

The memory observation is available headroom, rather than installed RAM:

• Linux takes the minimum of available host memory, finite process address-space and data-segment headroom, and readable cgroup memory headroom at the leaf and every ancestor.
• macOS uses native Mach free plus inactive pages, capped at half physical RAM, and applies finite process memory limits when available.
• Windows uses native GlobalMemoryStatusEx physical, available current-process commit, and virtual-address headroom.

A missing observation uses a 1 GiB memory reference or five million lookup operations per second. A readable exhausted-memory limit stays zero and denies calculation reservations. Missing host information does not discard readable process or container ceilings. If a finite process or container constraint is known but its current usage cannot be read, headroom is conservatively zero. The platform snapshot cannot guarantee that another process will not consume memory after admission.

Resolved resource	Formula and independent bounds	Unknown observations
Geometry allocator budget	headroom / 4, clamped to 64–512 MiB	256 MiB
Geometry deadline	300 million / lookup rate seconds, clamped to 60–120 s	60 s
Aromaticity work	lookup rate × 10, clamped to 5–200 million units	50 million
Aromaticity deadline	10 s	10 s
Aggregate calculation reservations	minimum of headroom / 2 and 1,280 MiB	512 MiB

A faster throughput observation never reduces the historical 60-second geometry deadline. The independent wall-clock bound remains effective for cases whose cost is poorly represented by the lookup probe. Explicit development-client limits must still satisfy the 120-second and 512-MiB ceilings.

Reservations and cancellation

Geometry reserves its allocator budget plus 32 MiB of process overhead before launching a disposable worker. Waiting consumes the request deadline and dropping the asynchronous operation cancels waiting or the worker exchange. The worker allocator enforces charged live Rust allocation bytes; this is not a total-process memory cap, since stack and operating-system allocations are separate. All reservations release on return or cancellation. Reservation accounting is shared within one application process; separate ReShiki instances do not share a reservation counter.

Aromaticity reserves a checked estimate based on atoms, bonds and ring slots, plus 16 MiB of overhead. This admission estimate is not an in-process allocator cap. Cancellation and deadlines are checked before and after perception and periodically during charged work. Failure returns an error without a partially perceived graph. These reservations cover geometry and aromaticity; they do not account for every application or operating-system allocation.

Structural and chemistry constraints

Geometry retains the historical 256-MiB / 512-atom admission anchor. Below that anchor it permits one original atom per 512 KiB of allocator budget; above it, pair-table growth requires one additional original atom per 2 MiB. Original bonds are limited to four times that atom envelope; all-atom coordinates are limited to eight times it. Independent structural ceilings remain 640 original atoms, 4,096 original bonds and 4,096 coordinates, including temporary hydrogens. Thus the 64/256/512-MiB budgets admit at most 128/512/640 original atoms. Eligibility does not guarantee parameter coverage, embedding success, convergence or completion within the worker deadline. The worker independently validates the requested envelope against its heap budget and the structural ceilings.

The framed protocol accepts a request without the new optional capacity field. Such a legacy request retains the 512-atom and 2,048-bond envelope, intersected with its heap-derived envelope. Normal clients relaunch their own executable; development clients must use a worker with the matching protocol schema. This compatibility does not allow a new request to an old worker.

Aromaticity retains the default model’s six-ring combination limit and its two-ring limit for fused systems larger than 300 rings. These are chemistry rules, independent of the operational work budget. Ring-slot and ring-neighbor ceilings also remain in force.

Connected fused-ring enumeration

Singleton rings and adjacent pairs are evaluated directly. Larger combinations are streamed in the former exhaustive lexicographic order of fused DFS positions. A sorted prefix may be disconnected: the path 0–2–1 must still emit [0,1,2]. A bounded 0–1 breadth-first search charges selected vertices zero and future vertices one, excluding already skipped positions. It prunes only when a selected vertex cannot be reached with the remaining slots or too few future vertices can be reached. Every complete connected combination therefore remains eligible. No full collection of combinations or exhaustive fallback is stored.

The existing stop condition runs only between complete combination-size levels. Dense systems can still require combinatorial work and fail the operational budget. Successful runs preserve atom and bond markings and aromatic-ring counts; work-limit success and failure boundaries intentionally improve with the new enumeration.

Validation and limits of the envelope

The issue’s 37-repeat fused-ring input has 452 atoms and 150 rings. The previous enumeration exceeded its 50-million-unit budget; the new enumeration completes with 385,972 charged units. Exhaustive connected-subset checks cover every labeled graph through five vertices and compare output order. The independent aromaticity reference check also covers 31,019 cases, including 17,307 transformed graphs. The six-ring and greater-than-300-ring rules remain unchanged.

Actual seeded geometry probes exercise Generate, Evaluate and pinned Relax at 513 and 640 original carbon atoms under MMFF94, MMFF94s and UFF, checking finite coordinates, hydrogen mappings, energy agreement, finite analytic gradients and unchanged pins. At 640 original carbons, UFF generation used about 421 million charged live allocation bytes under the 512-MiB worker budget. Larger 768- and 1,024-carbon probes exhausted that budget safely, so they are not admitted. The conservative 640-atom ceiling retains headroom; it is not unlimited scaling with installed RAM. The new framed 511/512 controls retain the baseline’s exact seeded coordinates, energy and hydrogen mappings.

Development builds optimize cosmolkit-core at level 3, matching the numerical test profile. A 513-carbon, 1,541-coordinate ordinary 500-iteration calculation still exceeded a resolved 60-second deadline; an explicit, valid 120-second developer request completed in approximately 69 seconds with converged=false. The retained calculated carbon seed subsequently completed an ordinary request in approximately 73 seconds under its resolved 75-second deadline. A separate 513-original-atom alkane with 171 carbons and 342 explicitly drawn hydrogens completed the ordinary 500-iteration request in approximately 9 seconds under its resolved 60-second deadline, returning 515 all-atom coordinates. Both ordinary controls preserved the source drawing and reported converged=false. Thus eligibility and a complete result are distinct from convergence. Timed-out work returns no geometry to apply. These measurements establish bounded capability on the tested machine, not a timing guarantee for every eligible molecule or machine. Cross-platform resource probes require runtime validation on their respective platforms.