Add mechanical stuck-loop guard for agent loops

[JE-Chen]

Computer-use agent batch (from multi-agent web research; verified absent). Full layers + tests + EN/Zh v46 docs + README.

Feature (utils/loop_guard, pure-stdlib)

• LoopGuard watches the (tool, args, result) stream and flags three stuck-loop patterns: repeat (same call N times in a row), ping_pong (A-B-A-B alternation), no_op (observation digest never changes), escalating ok→warn→critical by run length (configurable thresholds, bounded window). digest_result makes a stable short hash of a screenshot/observation (bytes or JSON-able).
• Why: the dominant computer-use failure mode is an agent burning budget on a no-effect loop the model can't see; a step/time budget can't distinguish a productive loop from a stuck one, and trajectory eval is offline. This is the missing in-loop, mechanical detector.
• Executor AC_loop_guard_observe / AC_loop_guard_reset (module-level default guard); MCP ac_*; Builder under Agent.

Verification

• 10 tests pass (repeat escalation ok/warn/critical, repeat count, distinct-args not-repeat, ping_pong, no_op on unchanged digest, reset, digest stability for bytes/dict, executor round-trip, wiring, facade); ruff clean; radon no CC≥C; bandit clean; PySide6-free.
• Stacked on coordinate-space (Add coordinate-space mapping (model grid <-> physical pixels) #253); rebased cleanly onto dev.

[codacy-production]

Up to standards ✅

🟢 Issues 0 issues

Results:
0 new issues

View in Codacy

🟢 Metrics 58 complexity · 0 duplication

Metric	Results
Complexity	58
Duplication	0

View in Codacy

_{NEW Get contextual insights on your PRs based on Codacy's metrics, along with PR and Jira context, without leaving GitHub. Enable AI reviewer
TIP This summary will be updated as you push new changes.}

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud