great_cto

Verified acceptance — the system stops pretending it succeeded

A continuation of the previous release's theme, informed by the "code clean-up
agents" argument that the bottleneck has moved from generation speed to
verified, minimal, reviewable change — not faster auto-merge loops. Three
additions, each built by composing pieces already in the repo; none adds a
runtime dependency, new tracking, or recurring cost.

Edit-scope enforced at write time, not after

• New PreToolUse hook (Edit|Write|MultiEdit) turns an IMPL-BRIEF's declared
  scope into a hard constraint on the write, reusing the existing checkScope.
  A denylisted file → hard deny (exit 2), always; a file on neither list →
  advisory by default (scope creep surfaced, not blocked, because incomplete
  allowlists otherwise train people to disable the hook);
  GREAT_CTO_ENFORCE_EDIT_SCOPE=block upgrades it; no active brief → allowed.
• The active brief is located via the .great_cto/active-brief pointer that
  senior-dev writes on task claim and clears on close/block. A stale pointer can
  only ever block a denylisted path (fail-safe); GREAT_CTO_DISABLE_EDIT_SCOPE=1
  is the escape hatch. 12 tests + a verified stdin→deny/allow e2e.

A stopped agent emits a handoff package, not a bare status flip

• scripts/handoff-package.mjs <reason> <feature> assembles one reviewable
  evidence bundle when an agent stops (blocked, spec-objection, budget cap):
  the diff + changed files, what ran and how it came out, the attempt history,
  what it cost, and why it stopped. Reads only local logs already on disk
  (agent-writes, cost-history, verdicts, git) — no new tracking, no ADR. Reports
  honestly: unmeasured tests are "unverified", a stop before any write says so
  rather than fabricating a diff. Wired into senior-dev's blocked +
  spec-objection paths. 8 tests.

Acceptance-oriented metrics on the board

• cost_per_accepted (window AI cost ÷ gate-APPROVED verdicts) and
  rework_rounds (sent-back review outcomes), derived from verdict logs and
  surfaced as two board tiles. cost_per_accepted is null when nothing was
  accepted in the window — dividing a real cost by zero approvals would invent a
  number, same discipline as the null-score oracle. Deliberately two of the five
  metrics the literature lists, not all five: a solo project's denominator is
  tiny and the rest would be noise until volume justifies them. 8 tests.

Full suite green across board / top-level / hooks / lib (107 / 207 / 168 / 368).