Keep requirement ownership with the user and the wording with the orchestrator

[shinpr]

What changed

The user's wording stays in the orchestrator. requirement-analyzer received the full request and returned a six-category classification of that same wording. The orchestrator already holds it, and requirements_verbatim at the Design stop already comes from it. The analyzer is now invoked with a minimum evidence contract — the shortest verbatim outcome plus the shortest reason needed to interpret it — and returns repository evidence only. Evaluation requests, speculative ideas, and prescribed mechanisms are classified in requirement-convergence from the retained wording.

Scope evidence starts from the responsibility, not the paths. Discovery began at likely target paths, so an existing surface already owning the requested responsibility surfaced only if it happened to sit on one. Each retained boundary now records its current treatment alongside the consequence a user would observe. The depth gate and branch stop condition are unchanged.

The requirements stop has one output shape. Confirmed scope, Decision evidence, User decisions, and Workflow are defined once in requirement-convergence; the recipes add only what the shared shape cannot know. Separating settled boundaries from repository observations from open questions is what makes the decision informable. Only an explicit user answer moves an item into Confirmed scope.

Each stop selects its own confirmation form. The blanket directive routing every confirmation through a selectable-option prompt is removed. At the requirements stop the candidates would have to be authored by the orchestrator, which records the orchestrator's wording as the user's decision and contradicts the hearing's own pass condition. Stops keep their wait-for-confirmation rule; the sites whose candidate sets are enumerable outside the session keep their own directive.

Analysis-exposed product boundaries return to the requirements gate. Change detection fired only on a proposed change, so a user-facing responsibility surfaced by codebase or UI analysis entered design scope without reaching the user. It now returns only when leaving it or changing it would alter the confirmed outcome or an exclusion; technical choices continue to the design owners.

Ownership wording. The user owns product requirements and exclusions; the orchestrator owns convergence readiness, Structural Scale, and routing.

Reverse-engineering discovery is gated on a non-empty unit inventory. This reinstates the all-empty check, not the per-category one removed in #181: a unit with routes but no tests proceeds, a unit with no routes, tests, or public exports does not. unit_inventory is the completeness baseline for code-verifier, and three empty categories balance at inputCount: 0, so such a unit passed verification with no evidence examined and no limitation reported.

Handoff context is tested for consumption. llm-friendly-context now checks that every included item is consumed by the target action, alongside the existing sufficiency checks.

Version 0.27.0 to 0.27.1.

🤖 Generated with Claude Code