M12.4: retain and settle owned population intervals
Foundation / checks (deploy) (push) Successful in 5s
Foundation / checks (ecs) (push) Successful in 2m28s
Foundation / checks (tsan) (push) Failing after 18m22s
Foundation / checks (network) (push) Successful in 20m38s
Foundation / checks (client) (push) Failing after 21m6s
Foundation / postgres-tools (push) Canceled after 0s
Foundation / server-image (push) Canceled after 0s
Foundation / image (push) Canceled after 0s
Foundation / checks (postgres) (push) Canceled after 46m54s
Foundation / checks (checks) (push) Canceled after 47m3s

This commit is contained in:
dingo
2026-10-17 17:31:05 -04:00
parent 627411983f
commit 0a8962b57e
34 changed files with 1032 additions and 75 deletions
@@ -0,0 +1,71 @@
# ADR-0143: Owned bounded population intervals
## Status
Accepted for M12.4's opt-in engineering path, 2026-10-17. Extends ADR-0133,
0141 and 0142 without changing biology, controller decisions or World0 policy.
M12.4b's interval segment is implemented and validated. Biological migration and
the rest of M12.4 remain separate work; this does not close the entire gate.
## Context
Local confirmed-head pacing preserves debt while an owner runs, but loses its
unpublished interval on process loss. Inferring a replacement interval from elapsed
wall time would violate the accepted outage pause and could change controlled
biology. Profile 5 already supplies exact frozen-player/wildlife replay and bounded
publication. It needs a durable operational admission boundary.
## Decision
Admit one interval per region under its current lease and exact profile-5 head.
Bind the source commit/digest, begin/end ticks, controller version, admitting
holder and epoch. Intervals contain 1..64 unchanged 50 ms biological steps. Keep
admission records immutable. Identical retries return the same interval; different
end boundaries for one source conflict. New admissions require the exact current
head and no unsettled predecessor interval.
Write its immutable settlement with the checkpoint in the same transaction.
Normal profile-5 biological publication must consume precisely its admission;
manual conversion or profile exit waits while any interval is unsettled. Old
committed records remain recoverable and exactly retryable without fabricated
retroactive admission records. Generic publishers cannot bypass these guards.
A replacement fenced owner loads pending admission before selecting fresh work.
It advances from the verified source using the recorded controller version and
frozen controlled intentions, then publishes the exact trace/payload. Unknown
admission replies retain the same request until reconciliation; unknown checkpoint
replies retain their existing prepared bytes. Read-only resolution grants no
authority. A returning player waits for this bounded admitted interval, then the
host materializes before admitting player operations.
New profile entry and interval preparation require every controlled cell to be
fully suspended. Frozen intentions alone are insufficient: during the existing
200-tick linger period a cell is still vulnerable to predation, and its permanent
ending must stay on the existing profile-4 path. That path finishes linger and
records any death before entry; no new immunity or biological timer is introduced.
Earlier profile-5 traces remain verifiable with their original semantics and can
be materialized. Recovery reconciles retained owned corpses for either profile,
so an older post-commit/pre-ending crash cannot leave its account bound forever.
Collect one second of due time for routine unattended admissions, service at most
64 ticks per batch, and preserve debt until the confirmed head consumes it. A
short terminal-u64 remainder is admitted without waiting for 20 ticks. Report
pending admitted ticks separately from local due debt and checkpoint freshness.
Capacity pressure pauses admission; it cannot drop an admitted interval or cull
organisms. Unadmitted wall time remains paused across outage (ADR-0133).
## Alternatives
Keeping only a local deadline loses admitted work on restart. Reconstructing it
from database timestamps adds an unaccepted catch-up policy. A transaction per
tick violates the founding durability requirement. An unbounded queue conceals
resource commitments and makes returning-player latency unbounded.
## Consequences and limits
This adds a bounded operational journal and transaction guards, not a new time
integrator or ecological rule. Require duplicate/conflicting admissions, lost
replies, stale writers, process loss before/after admission and publication,
restart with zero new due time, and exact uninterrupted continuation comparisons.
Migration escrow, delayed boundary coupling, environmental ownership and multi-day
evidence remain separate M12 work. Production enablement stays off by default.