docs(protocol): v1.11.0 — evidence-first questions, unattended modes, trust-model hardening

Applied a 42-finding multi-agent review of the protocol across consistency,
security, efficiency, usability, coverage, cross-tool, parity, and
enforceability lenses. Highlights:

- Understand-before-acting: self-serve research before asking, recommended
  answer per question, weight-gated design dialogue
- Trust & Safety: scoped AGENTS.md-edit exception, expanded high-risk list
  (exfil channels, network-fetched code, CI/hook edits, scope expansion),
  forwarded third-party content carries no user authority, split
  source-priority rule into short sentences
- Standing corrections capped to preference/convention level
- New failure modes: UNATTENDED (headless/CI) and CONTEXT_LOSS (checkpointing)
- Startup: minimal task-triggered bootstrap path; step 4 dedup; steps 1-3
  runnable vs 4-5 standing
- Evolution: promotion/demotion counted from recorded outcomes, experiments/
  candidates ledger-tracked, slimmer outcome schema, scoped promotion row,
  gated routine updates to active capabilities, AGENTS.md row covers
  create+nested, memory-deletion exception per Failure Modes
- Template update: official-source-only fetch, no offline reconstruction,
  confirmation when safety sections change
- Maintenance: observable health-check trigger, session-data reconciliation,
  single pinning mechanism
- zh-CN: 7 parity fixes (modality drift, dropped qualifiers, 会话状态)

Author: yeasy