idkmesh

Issue #84 Factor-Isolation Closure

Date: 2026-08-29

Project-owner requirement

Select another open issue with no overlapping active work, solve it through a focused pull request, and merge only after current exact-head evidence passes.

Selection and scope

Issue #84 was unassigned after its Phase A PR merged. A fresh audit found no open PR, local branch, remote branch, or worktree targeting its remaining Phase B/C scope. A capability-rarity claim had in fact been posted nine seconds before this work’s claim and then merged as PR #262. The branch was rebased onto #262 and its duplicate capability implementation was discarded. This contribution covers only the independent lag, failure-shape, saturation, and coordination- cost work that #262 explicitly left open.

The remaining acceptance surface was bounded to factor isolation and stronger coordination-cost evidence. The work deliberately does not expand into real fleet validation or claim that one scheduling policy is universally optimal.

Implementation

randomness_lab.r2_factor_sweep adds five-seed sweeps for:

Each raw policy run retains performance, failures, churn recovery, locality, metadata probes, modeled messages/bytes, capability-directory operations, and scheduler state. An opt-in profile separately measures process CPU time and Python traced peak allocation on a recorded host.

Findings and decision

PR #262 establishes the capability-rarity result. This continuation finds that availability lag creates unreachable routing attempts, while load lag instead increases queue response time. The regional-outage effect is measurable but modest in this model. Near saturation, local two-choice latency diverges from the full-information oracle.

These results complete issue #84’s synthetic Phase B/C acceptance criteria and preserve negative regimes. Real packet measurements, distributed directory behavior, and hardware/fleet validation remain explicitly out of scope.

Community impact

Contributors can now inspect one raw, reproducible causal matrix instead of inferring causes from combined fresh/moderate/stale presets. Transparent cost assumptions make it easier to challenge or replace the model.

AI/tool provenance

Codex implemented the harness, tests, reference artifacts, and synthesis. Local verification passed the self-test, 22 focused R2 tests, 439 repository tests, and 25 interop tests (two optional-dependency skips). The deterministic matrix has SHA-256 988b002bdbbd28a92c34825381ccbe45f09384780494a3300f91457987f680fa; the host profile has SHA-256 0265c4cc971b7e4f8cbf0f8113a2850676378540715a4cbacc44f36b81cdc0ee. GitHub exact-head verification remains required before integration.