test(e2e): advance fixtures from tool receipts

[wstczyw]

Summary

• advance the direct-skill E2E fixture from observed tool receipts instead of HTTP request counts
• isolate native memory_consolidation traffic from the foreground scenario
• capture the launch job id and substitute it into the next shell command
• add a regression covering retries, auxiliary traffic, and failed receipt processing

Why

Native background requests and retries can arrive at the mock provider while a foreground skill scenario is running. Counting POST requests lets that traffic consume a shell step, causing the fixture to execute an unresolved <job-id> placeholder and produce misleading failures or hangs.

Tests

• node --test --test-name-pattern="durable tracked review fixture isolates memory traffic and retries|requires the codex CLI|durable tracked task and result readback" tests/e2e/codex-skills-e2e.test.mjs
• npm run lint -- --no-warn-ignored tests/e2e/codex-skills-e2e.test.mjs
• npm run typecheck
• npm run check:version-sync
• npm run check:changelog

This is test-only; production routing and memory behavior are unchanged.