M12.2: select and qualify exact cellular local cohorts
Foundation / checks (deploy) (push) Successful in 5s
Foundation / checks (ecs) (push) Successful in 2m19s
Foundation / postgres-tools (push) Canceled after 0s
Foundation / server-image (push) Canceled after 0s
Foundation / image (push) Canceled after 0s
Foundation / checks (tsan) (push) Canceled after 6m24s
Foundation / checks (network) (push) Canceled after 6m33s
Foundation / checks (postgres) (push) Canceled after 6m34s
Foundation / checks (checks) (push) Canceled after 6m44s
Foundation / checks (client) (push) Canceled after 6m44s

This commit is contained in:
dingo
2026-10-17 12:04:39 -04:00
parent bee57e5829
commit 2b074b9ae1
17 changed files with 762 additions and 149 deletions
@@ -0,0 +1,63 @@
# ADR-0139: Exact retained-member cellular local cohorts
## Status
Accepted for the bounded M12 cellular population implementation, 2026-10-17.
The maintainer delegates M12.2 decisions to the implementing agent, subject to
later revision. This decision does not enable a production selector, increase
capacity, qualify approximate aggregation or authorize deployment.
## Decision
Replace the rejected M4 averaged model with an exact local model for rules 8.
A cohort groups wildlife of one genome in one region, retaining every member's
identity, position, birth tick, growth, damage, reserves, reproduction state and
ancestry. There are no bins, fractional members, random materialization samples,
mean reserves or statistical parentage. The residual is identically zero.
`CellularPopulation` exclusively owns one `CellularLives` store. Sorted membership
tags partition that store into active individuals and cohort members; they are
not a second inventory. The read-only biological view conveys no spending rights.
All actions execute through the population owner. Controlled cells, including
offline/suspended cells, never collapse. Collapse/materialize operates on the
entire eligible wildlife genome group, never a caller-picked healthy subset.
New offspring initially have individual representation. Dead membership is
removed after the successful atomic biological step.
The same bounded wildlife controller and complete shared interaction kernel run
at the existing 50 ms quantum, in the existing life-ID order. Cohort membership
cannot affect intake contention, movement, hunting, budding, mutation, death,
inventory or ancestry. The host submits only controlled intentions; it cannot
supply a different controller for cohorts. No autonomous mutation is introduced.
Biological rules, existing codec versions and old goldens are unchanged.
Keep regional approximation disabled. Do not silently substitute the old M4
model under overload. Retain the existing hard limit of 256 living cells per
region; new admissions backpressure rather than deleting incumbents. This is a
structural ceiling, not a measured service capacity. The operational target is
completion within the 50 ms tick, with measured tail latency and headroom on the
actual deployment host required before enabling a production selector. An
overrun must leave admitted ticks pending, not skip ecology.
Qualification is declared before results: reference seeds 0/1, held-out 2/3,
rich/scarce finite food, conversion oscillation every 1/16/64 ticks, paired full
canonical biology equality every tick. Zero demographic, biomass, diversity,
ancestry or inventory discrepancy is allowed, including false local extinctions.
Report conversion coverage and full owner costs, including active fallbacks at
16/64/128/256 lives. Bin sensitivity is inapplicable: no state is binned. Do not
infer unobserved predation, twins or long-term stability from passing cases.
## Consequences and limits
This resolves the unsafe averaging conflict for an exact local implementation
and permits work on its durable representation. It deliberately trades memory
and computation for fidelity. Cohorts currently provide ownership organization,
not population compression or a demonstrated speedup. The M12.2 target-host cost
gate and scalable approximate model remain open; do not mark the whole gate
passed solely on this decision. Production enabling and higher capacity require
new measured evidence. Future approximation requires a separate version/ADR and
paired evidence; it must not reinterpret these exact members.
The module lives in portable `evo::lifecycle`, independent of renderer, network,
database and clocks. The foundation proof and legacy ecology experiments retain
their existing boundaries. No arbitrary third-party dependency is added.
+2
View File
@@ -155,3 +155,5 @@ design. No unadopted candidate is a link dependency of the proof.
- [ADR-0087: Dev session route](0087-dev-session-route.md) - evolution-dev only: void.aeternalisgames.com:7400 via a dedicated internal MetalLB VIP to the region pod, TLS end to end, CIDR-limited, operator-opened windows enforced by evo-server; committed disabled with fail-closed policy tests.
- [ADR-0086: Windows client package](0086-windows-client-package.md) - static-runtime Windows x64 ZIP with explicit allowlist, licence texts, build identity, SHA256 manifest, credential guards and deterministic layout; local packaging until a Windows runner exists.
- [ADR-0085: Rebindable engineering controls](0085-rebindable-engineering-controls.md) - SDL-free action/HID-key bindings, conflict validation, versioned all-or-nothing controls file, binding capture before toolkit/world.
- [ADR-0139: Exact cellular local cohorts](0139-exact-cellular-local-cohorts.md)