daemon start stopped crying wolf about stale log lines

The previous goalpost (PR #2330 / LIVE-25) stopped an unrelated pytest run from SIGTERM-ing a healthy daemon via the shared HTTP port reclaim path. With the daemon actually staying up, soak testing immediately surfaced a different false alarm: synlynk daemon start printing ⚠ Failed to start daemon: SystemExit: 143 and synlynk watch is also running, then synlynk daemon status confirming a healthy process seconds later.

What changed

Filed as gh:#2333. Root cause was two compounding bugs in WatchDaemon.start() (shared by SynlynkDaemon), plus a third stale-state check on the same start path:

1. _await_child_pidfile() used a flat 2.0s cutoff. Under load the re-exec child often published its pidfile just after that window, so start() took the failure branch even though the child came up fine.
2. The failure branch tailed the entire append-only logfile. A quiet successful spawn writes nothing new, so the last non-blank line from an old diagnostic session (here: leftover SystemExit: 143 from the LIVE-25 investigation) was misreported as today's error.
3. SynlynkDaemon._prepare_start() treated watch.pid existence as "watch is running". A dead pidfile triggered the dual-poller warning with zero live watch processes.

The fix. Pidfile wait now polls with exponential backoff up to a 10s ceiling. Before spawning, start() records the logfile byte offset and only reads lines appended after that watermark; if nothing new appeared, it reports an honest readiness-timeout message. The watch check now requires a live PID (or live lock owner), matching _health() liveness rules.

Tests. New file tests/test_daemon_start_false_failure_2333.py covers: stale log not misreported on a silent failed wait; backoff past 2.0s; stale watch pidfile ignored; live watch pidfile still warns; only newly appended log bytes reported on a real failure. Related suites (test_daemon_liveness_1572, test_control_plane_daemon_health, test_daemon_token_refresh) stay green: 47 passed.

Goalpost

Daemon start diagnostics now match reality: a late-but-successful child is waited for, stale log tails stay in the past, and a leftover watch.pid no longer claims a concurrent poller. Next soak pass can trust Failed to start as a real signal while the autonomy-pass loop from the prior PR cluster keeps running unattended.