0.84.0-alpha · core 2026-10-22

The published context policy and the enforced gate disagreed, in BOTH directions. context:5 skipped the Context stage for an experimental probe under Auto; the helper skipped it for a single direction too. So a worker that followed the policy was refused by the gate with no redirect (invalid transition 'setup' -> 'shape'), and a worker that followed the gate ran a stage the policy said to skip. The two answers came from two owners of one rule. The rule is now ONE fact in stages.yaml (routes.stage_policies), rendered into two faces that cannot disagree: the decision table the LLM reads in stages/context.md, and the machine-readable block the helper parses out of references/transitions.md. A new parity test reads the table the way the LLM reads it — first matching row wins — and compares its answer with the helper's for all 48 (quality, exploration, review mode) combinations.
A legacy appetite state carried a third copy of the rule. A state with only appetite: Lean and no knobs skipped the Context stage; the documented mapping says Lean resolves as {production, 2}, and a production run with two directions runs the stage. The mapping is the owner now, and the third copy is gone.

The Context stage is skipped when its answer changes no consequential decision AND nobody reviews it. The rule is now: Auto review mode AND (quality is experimental OR exploration_count is 1). The Context stage is a spike — it answers one question, "which strategic approach?" — and a spike is worth running only when it reduces uncertainty that will change a consequential decision, and when its answer has a consumer. A probe will not ship as-is, so its strategic choice is not acted on as a product decision; one direction is "trivial changes only, by explicit choice", so there is no divergence to explore; and a gated review mode has a person reading the output, so cutting the stage there removes their input.