v1.12.0 — Failure-mode and delegation refinements

Applies the five single-vote findings from the v1.11.0 multi-agent review, each reconciled with the dissenting verifier's objections rather than taken as proposed.

• BROKEN_ENV failure mode: when tools, dependencies, or validation commands cannot run — report the exact failing command, propose the smallest fix, ask before extended environment repair, deliver as explicitly not validated only with user consent, and record the working setup in workflows/.
• Delegation contract (Own the outcome): every subagent gets a self-contained brief — goal, scope, relevant paths, constraints, expected output — and reports results back to the delegating session; shared .agents/ writes follow CONCURRENT_WRITES.
• Memory scoping in monorepos: one .agents/ per protocol install, beside the AGENTS.md that carries it; nested AGENTS.md instruction files do not create additional layers. (The reviewed per-scope-.agents/ variant was rejected: it would fragment standing corrections, changelogs, and health-check state.)
• Secret-read guard: do not read credential files or secret stores unless the task requires it; never echo secret values into responses, command-line arguments, or error output.
• Bootstrap step 3c: the deterministic created-at-bootstrap set stays explicit; the remaining Directory Layout memory files (decisions.md, gotchas.md, patterns.md, project-map.md, secret-requirements.md) are declared created-on-first-write, resolving the layout ambiguity without empty placeholder files.

Both templates remain line-for-line parallel (302 lines).

Full diff: v1.11.0...v1.12.0