Fix silent live session stalls

[saleh-mir]

Summary

• bound synchronous Redis connection and read operations so a worker cannot wait indefinitely
• keep a running worker active during transient Redis connectivity failures, with rate-limited diagnostics
• preserve real background-thread tracebacks and guarantee failed sessions terminate even if secondary reporting fails
• report child-process exit codes before cleanup so abnormal exits remain diagnosable
• retain failed Redis worker-marker removals for retry, without removing a marker that belongs to a newer worker using the same client ID
• stop deferred cleanup after the first failed Redis call in each cycle so an outage cannot stack one-second waits while holding the worker lock

Root cause

PR #615's original cleanup path removed a finished worker from local tracking before removing its Redis marker. If that Redis call timed out, the stale marker remained permanently and could make a dead session appear active forever.

Verification

• PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 /Users/salehmir/miniconda3/envs/jesse3.12/bin/python -m pytest -q — 624 passed
• /Users/salehmir/miniconda3/envs/jesse3.12/bin/pyrefly check jesse/services/multiprocessing.py --search-path . --conda-environment jesse3.12 — 0 errors
• python -m compileall -q jesse/services/multiprocessing.py
• git diff --check
• failure-path harness covering Redis timeout/recovery, full-reset recovery, client-ID reuse protection, rate-limited diagnostics, and one-timeout-per-cycle behavior

Notes

This supersedes #615 because its GitHub pull-request head ref became detached from the live source branch and remained pinned to the earlier commit.