Skip to content

Decision record · ADR

ADR-0032: Classify time controls by observable ownership

Status: accepted | Date: 2026-08-04

References: RFC-0009, RFC-0003, RFC-0004, RFC-0005, ADR-0015, RFC-0006

A remote readiness incident showed that a fixed 250-millisecond HTTP-attempt cap could reject every valid response even while the operator’s readiness budget still had almost an hour remaining. A production-path audit found the same structural ambiguity in shorter connection caps, prefix-cache reset, profiler control, SSH liveness options, adapter operations, profiler finalization, and cleanup. Some of those durations can change whether a valid experiment succeeds; others only bound cleanup or control responsiveness; still others duplicate configuration already owned by OpenSSH or another external tool.

The existing operation-owned deadline model in ADR-0015 prevents nested durations from extending an operation, but it does not decide when a subordinate cap is a private implementation policy, when it is an operator-visible execution fact, or when InferLab must defer to another configuration authority. Treating every duration alike would either preserve hidden correctness limits or flood ordinary workflows with low-level knobs.

The decision must retain one resolved execution authority, framework-neutral Rust lifecycle ownership, framework-specific integration ownership, bounded and evidenced cleanup, and deliberately unbounded capture readiness. It must not introduce a global timeout policy object, a workflow scheduler, or source-pattern policing of numeric constants.

We will classify each runtime time control by the observable decision it owns before deciding its configuration surface.

An operation budget belongs to the concrete workflow whose terminal outcome it determines. A subordinate attempt that has no retry or alternative after expiry consumes that owner’s remaining budget instead of acquiring a shorter private deadline. A retrying or deliberately unbounded workflow may use a bounded attempt to regain lifecycle control, but an environment-sensitive bound that can repeatedly exclude valid local or remote responses becomes a typed, resolved workflow setting with a documented default and recorded effective value.

Cleanup grace, polling and retry cadence, progress timing, and other control-plane responsiveness policies remain domain-owned product policies while they do not replace a business terminal decision. Values that affect cleanup verification remain explicit in cleanup evidence. External transport configuration remains with its owning tool; InferLab uses resolved operation bounds for its own lifecycle and does not mirror or silently override operator-owned OpenSSH liveness policy.

The server, measurement, profiler, adapter, container, and SSH domains retain separate policy ownership. Shared code may carry remaining-budget mechanics, but no global timeout configuration object or public collection of implementation constants is introduced.

Operators gain control over latency-sensitive attempt bounds that vary with placement, while ordinary workflows do not acquire settings for polling, presentation, or routine cleanup internals. One-shot requests can no longer fail under a hidden duration shorter than their declared owner budget, and remote transport settings retain one established authority.

The control plane must carry new effective fields and evidence only at boundaries where time policy is genuinely operator-visible. Some existing fixed values remain valid product policy, so audits require semantic classification rather than a mechanical constant ban. Defaults that affect correctness require specification and compatibility review instead of being changed as private tuning.

Removing OpenSSH overrides requires bounded or interruptible ownership around InferLab commands rather than relying on SSH options as an accidental supervisor. Profiler finalization and similar multi-action cleanup paths require one enclosing cleanup budget where they claim one finalization deadline. These migrations are separate domain outcomes and can be delivered incrementally under RFC-0009.

Classify time controls by observable ownership (accepted)

Section titled “Classify time controls by observable ownership (accepted)”
  • Pros: Keeps configuration proportional to operator-visible choices, Removes hidden premature failure without turning cleanup and cadence into public API, Preserves OpenSSH and other external tools as authorities for their native transport policy
  • Cons: Requires a semantic audit and domain-specific migrations instead of one mechanical rewrite, Adds resolved fields and evidence at the few latency-sensitive workflow boundaries

Expose every runtime duration as configuration (rejected)

Section titled “Expose every runtime duration as configuration (rejected)”
  • Pros: Makes every duration directly adjustable
  • Cons: Leaks polling, presentation, cleanup, and implementation mechanics into ordinary workspace configuration, Creates multiple knobs for facts already owned by external tools or enclosing operation budgets
  • Rejected because: Configurability is not ownership; exposing all durations would make the product surface larger while preserving ambiguous semantics.

Keep every duration as a domain-local constant (rejected)

Section titled “Keep every duration as a domain-local constant (rejected)”
  • Pros: Keeps the public schema small
  • Cons: Lets machine and placement latency silently determine whether otherwise valid workflows can succeed, Duplicates standard transport configuration and hides compatibility-sensitive defaults in implementation
  • Rejected because: The remote readiness incident proves that some attempt bounds are real operator policy, not private tuning.