Stop the whole fleet stampeding one advisory lock at boot

[maximusunc]

All 23 DB-dependent containers start the instant Postgres reports healthy, and each one took a fleet-wide advisory lock to run the idempotent schema upgrades. On any volume created since those indexes landed in init_db.sql the upgrades are a no-op, but every container still queued for the lock, and whoever lost the queue blew postgres_pool_timeout and logged a PoolTimeout traceback on an otherwise healthy startup.

Add a pre-flight catalog check outside the lock, so the already-applied case costs one indexed read and no serialization, and retry the upgrade with the same backoff the query paths use instead of burning a single 5s attempt when the DB blips during those first seconds.

Also raise max_connections 200 -> 300. Both compose.yml and config.py document that the sum of per-container pool maxima must stay under it, but 22 workers at 10 plus the server at 30 is 250, so the invariant was already violated.

Claude-Session: https://claude.ai/code/session_01PQeAtdPFUgZ1BTB3UB6oXc

[codecov]

Codecov Report

❌ Patch coverage is 93.33333% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 54.06%. Comparing base (667465e) to head (6dd0ffa).
⚠️ Report is 18 commits behind head on main.

Files with missing lines	Patch %	Lines
shepherd_utils/db.py	93.33%	0 Missing and 1 partial ⚠️

Files with missing lines	Coverage Δ
shepherd_utils/config.py	95.32% <ø> (ø)
shepherd_utils/db.py	64.08% <93.33%> (+0.53%)	⬆️

... and 2 files with indirect coverage changes

---

Continue to review full report in Codecov by Harness.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 5a30daf...6dd0ffa. Read the comment docs.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.