[icesixgod]

Summary

Fix intermittent Windows failures in the task read-budget regression. Session discovery sorts by modification time; tied timestamps can put the requested task first, so the resolver legitimately returns before probing unrelated metadata. The old test assumed the unrelated file was always visited.

Exercise both discovery orders explicitly: unrelated-first must probe with the default byte limit, while selected-first must return without reading unrelated metadata. Keep the budget assertions strict and remove dependence on filesystem timestamp resolution.

Privacy impact

Test-only change. Runtime session discovery, byte budgets, returned task data, redaction, and consent behavior are unchanged.

Validation

• uv run ruff format --check .
• uv run ruff check .
• Reproduced the original assertion failure using equal file modification timestamps
• Focused read-limit regression suite: 29 passed, including both discovery orders
• uv run mypy
• Backend suite with coverage and resource-leak errors enabled: 540 passed, 1 platform skip; 91.83% coverage
• Synthetic logs only; no private data added

Validate all GitHub CI jobs before merging and verify the subsequent main-branch push run as well.