docs: audit gaps between task execution and persistent intent

[kai-linux]

Summary

Document an evidence-backed assessment of Agent-OS against the operator goal of persistent intent fulfillment across coding and non-coding work. Identify seven gaps, distinguish existing foundations from proposed behavior, and define staged delivery with concrete acceptance scenarios.

Correct the roadmap readiness claim and link the audit. This is documentation only: no runtime behavior, worker configuration, live objective, or scheduling changes.

Evidence and Review

• Audited source at 94fd9b3, retained runtime aggregates, and the merged [task] In the Readme, change the title from Agent OS to Agent-OS #355 / docs: update README title to Agent-OS #356 case.
• Distinguished historical symptoms from current protections and avoided treating task completion statistics as verified user outcomes.
• Reviewed staged diff and all local document links; no raw runtime logs or credentials included.

Validation

• python3 -m pytest tests/ -q: 741 passed.
• 12 local document links resolve.
• git diff --cached --check: passed.
• python3 bin/secret_scan.py --staged: passed.

The proposed behavioral acceptance suite and non-coding pilots are not implemented or claimed to pass by this PR.

[kai-linux]

Self-review complete. Checked the source references against the audit baseline, re-queried #355 and merged PR #356, and verified the metric aggregates. The document distinguishes observed behavior, historical samples, existing mitigations, and proposed capabilities. The two-file diff contains no runtime changes or raw operational data; no unsupported claim of an implemented acceptance suite. No findings in this documentation diff. Residual limitation: live cross-domain behavior and recovery still require the proposed end-to-end evaluations.