Agentic Swarm Simulator — P1 trajectory audit

newcomer: AGENT LIMIT. Step 2 reached the guide (HTTP 200), but quoted the navigation label ‘Read the guide’. It left or looped and stopped at step 6. Receipt: BEFORE.json, role newcomer, steps 2–6.
phone: AGENT LIMIT. Step 2 reached the guide (HTTP 200), but quoted the navigation label ‘Read the guide’. It left or looped and stopped at step 4. Receipt: BEFORE.json, role phone, steps 2–4.
buyer: AGENT LIMIT. Step 2 reached the guide (HTTP 200), but quoted the navigation label ‘Read the guide’. It left or looped and stopped at step 6. Receipt: BEFORE.json, role buyer, steps 2–6.
reader: AGENT LIMIT. Step 2 reached the guide (HTTP 200), but quoted the navigation label ‘Read the guide’. It left or looped and stopped at step 6. Receipt: BEFORE.json, role reader, steps 2–6.
developer: AGENT LIMIT. Step 2 reached the guide (HTTP 200), but quoted the navigation label ‘Read the guide’. It left or looped and stopped at step 6. Receipt: BEFORE.json, role developer, steps 2–6.

Shared harness problem: the rejected short quote was silently discarded; the next actor prompt asserted non-completion although no judge had run. The completion reader also mixed navigation and content. The local repair exposes content separately and records/explains rejected quotes. The minimum quote length, independent visual judge, content guard, task rail, app pages and step budgets are unchanged. Regression: original code fails (regression-red.txt), repaired code passes (regression-restored-green.txt); 42 existing swarm/content tests pass (regression-green.txt).

No app bug is confirmed in these five traces. No fabricated app failing tests are supplied. Selector, timeout, auth and viewport failure were not observed. Readability and real-human success remain unmeasured. All five original paths are labelled agent limit: each reached the guide at step 2, quoted a navigation label rather than result sentences, then wandered or exhausted its moves. A shared harness defect silently rejected that quote without telling the actor why; the local repair adds explicit rejection feedback and a separate page-content field. The test fails on the original harness and passes with the repair; 42 existing swarm/content checks pass. The phone path broke off at step 4 due to provider HTTP 502; the other four ended stuck. Its unknown-usage reservation remains held, and no retry run was made. The one allowed rerun kept task, roles, seeds, viewports, step limits, judge prompt and completion guards fixed: 0/5 tasks completed, 0/5 screen-confirmed, 1 visual-judge calls; 5/5 cases show no completion improvement. Super VL still rejects actual guide instructions as link-only; its existing-content exclusion also conflicts with a reading task. This remaining evaluation limitation is not an app bug. No target-app bug is confirmed and no app failing test was invented. Rerun accounted USD 0.04053384 including USD 0.03845760 held for unknown usage (known rerun token estimate USD 0.00207624); total P1 accounted USD 0.04564008, below the USD 0.20 stop (not a provider invoice). Receipts: TRAJECTORY-AUDIT.json, BEFORE.json, result.json, SWARM-RESULT.json and VERIFICATION.json. No Chrome or deploy.