fix: log the cause of a fatal error instead of dropping it

[DamengRandom]

Closes #30

The change

src/index.ts:48 passed the error as pino's second positional argument. Pino reads a leading string as a printf format string, so err was consumed as an interpolation value for a placeholder that does not exist and never reached the log — every crash of every job printed Fatal system error ⚠️❌⚠️❌: and nothing else.

-  logger.error('Fatal system error ⚠️❌⚠️❌: ', err)
+  logger.error({ err }, 'Fatal system error ⚠️❌⚠️❌')

This is the form already used at every other error call site in the repo. It was the only outlier.

The test

src/fatal-error.test.ts spawns the real CLI with DATABASE_URL pointed at a closed port, so initDb() rejects and main()'s catch handler runs for real, then asserts the cause survives into the output. No database, network or credentials needed — port 1 on loopback refuses immediately.

Verified it fails on the unfixed tree and passes on the fixed one:

### WITHOUT fix:
# pass 0
# fail 1
### WITH fix:
# pass 1
# fail 0

Checks

# tests 76
# pass 76
# fail 0

pnpm tsc and pnpm format:check both clean.

Observed vs reasoned

Observed by execution: the pino drop, with both call forms run against the same Error object in one process — one printed type, message and full stack, the other printed nothing. The end-to-end CLI run producing the bare line and exit code 1. The identity of the discarded error (Error: connect ECONNREFUSED). The test failing before the change and passing after. The full suite, tsc and format:check after the change.

Reasoned, not executed: that the same handler also swallows a createLlm() failure thrown out of runDailyJobs — that path needs an interactive daily-kpi run this scan could not perform. The fix does not depend on it; the initDb() path is covered by the test.

🤖 Generated with Claude Code

https://claude.ai/code/session_01YA7F47ZTqxswS4qy34VNdG

---

Generated by Claude Code