feature-intakelisted

# Feature Intake — Risk Lane Classifier & Router Every implementation prompt passes through this gate before code changes. The output is a **lane** (how much proof/ceremony) and a **confidence** (whether a human is asked), recorded in `specs/<slug>/SUMMARY.md` and used to route the rest of the workflow. > **The human does not classify risk. The harness does.** > Two independent axes: > - **Lane** scales with **RISK** → how much proof and ceremony. > - **Confidence/ambiguity** scales with **UNCERTAINTY** → whether to pause for a human. > > A high-risk-but-unambiguous task runs autonomously through heavy proof. > A low-risk-but-ambiguous task still pauses. Never ask a human "is this risky?" — answer > that mechanically. Ask a human only "did I understand you?" and "may I cross this boundary?" <HARD-GATE> Do NOT edit files, scaffold, or dispatch implementation work until intake is complete: the lane is assigned, the intake statement is emitted, and `Lane:`/`Confidence:` are written to `specs/<slug>/SUMMARY.md` (shape: `templates/SUMMARY.template.md`). A hard gate (see below) forces `high-risk` and cannot be self-downgraded — only a human narrowing scope may lower it. The orchestrator MUST write a `Lane:` line: `.claude/hooks/risk-corroboration.sh` blocks a commit whose diff trips a hard gate while the declared lane is below `high-risk`. (A *missing* lane only warns — fail-open — unless `RISK_CORROBORATION_STRICT=1` is set, so write the lane rather than rely on the hook.) </H