fix(chart): let every component finish starting before liveness judges it

[TheMeinerLP]

Found in production, minutes after the 0.6.0 rollout.

exitCode 137 (SIGKILL), reason Error
Liveness probe failed: connect: connection refused
container startedAt 10:58:03 — "Link service listening" at 10:58:58

The process was not crashing. It was still importing, and its own liveness probe killed it: initialDelaySeconds 5 + periodSeconds 20 x failureThreshold 3 = 65 seconds, against 55 seconds to reach listening.

The time is import time, not work — 0.6.0 adds the OpenTelemetry SDK and its exporters, which is a large import graph, and link has 200m of CPU to do it with. worker survived because it has 4 CPUs and a startupProbe, which it has had since the model download made the same point more loudly. bot and link never got one, because until now nothing they did at import took long enough to notice.

Both get one now. Liveness and readiness are not evaluated at all until a startupProbe succeeds, which is exactly the instrument for "slow to start" versus "hung".

Five minutes of budget against the 55 seconds observed. Being generous costs a later notice for a process genuinely wedged at startup; being tight costs a crash loop on a cold node. Only one of those is a working deployment.

The regression guard matters as much as the fix

.github/workflows/chart.yml now asserts that every component has a startupProbe and that its budget exceeds its liveness budget — because a startupProbe shorter than liveness buys nothing while looking like it does. Rendered:

	startup budget	liveness budget
bot	300 s	70 s
link	300 s	65 s
worker	1200 s	60 s

Counter-checked: removing link's stanza from the template makes the assertion fail.

Also documented in docs/operations.md §5, since the symptom — 137 with connection refused — reads like a crash and is not one.

Current production state

Both link pods are running; one restarted once and has been stable since. This is not an outage, it is fragility: any restart on a cold node can exceed the budget again, and bot is exposed too at 1 CPU.