[ViperJuice]

Detailed plan for slice C3 of see #235: the last open slice of the test-hermeticity umbrella. Plan only; src/ and tests/ unchanged.

Plan: .consiliency/plans/detailed-235-c3-clock-injection-20260923-1229.md

Design in brief

• ClientManager(clock=...) is injected through the constructor; the default is time.time. Seven of the 18 time.time() call sites in manager.py join the seam: every site that stamps or compares a request's started_at / last_heartbeat / last_activity_at. The other 11 are reported timestamps or IDs that are never compared with a request stamp, and they stay on wall time.
• The idle-poll slice (min(idle_timeout_s, 1.0)) becomes a named constant tests can shrink. A fake clock that counts reads lets each test step time exactly once per idle check, so no test waits in real time and none asserts an upper bound on wall-clock time.
• Monotonic is out of scope, as a named follow-up. These stamps are epoch values used outside the manager (cli.py:1478/1549, handlers.py:6335), so a monotonic default would make gateway.list_pending show 1970.

Verified independently by the integrator

Applied the plan's diff and test bodies to a scratch worktree off current main, adding the three test imports the plan lists in its changes table:

• tests/test_client_manager.py: 253 passed; ruff check, ruff format --check and mypy src/ clean.
• The three heartbeat tests dropped from 0.50 s / 0.41 s / 0.20 s of real waiting to 0.01 s each.
• Mutation: moving the idle-timeout comparison back to time.time() made exactly the three named tests fail. Restored byte-identical.
• 30 of 30 flake runs passed.

🤖 Generated with Claude Code

https://claude.ai/code/session_013G8iK4noj1iNbSTqJVFEX6

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}