fix(crew): stop startup validation logging a false-positive error on healthy boots

[abhishekmishragithub]

Problem

The crew server dry-runs your setup_handler at startup to catch node __init__ / import / missing-env / graph errors before accepting traffic. But the canonical handler ends with await session.start(), and the dry-run intentionally skips session.initialize(), so start() hit raise ValueError("Session not initialized") on every healthy boot:

ERROR | Startup validation failed — pod will not accept sessions. ValueError: Session not initialized

It also left the pod _ready = False for perfectly correct code. The docs had to tell users to ignore this scary error (quickstart "Expected boot output" note).

Fix

Add a dry-run mode to CrewSession:

• _dry_run is set only by the server's validator (real sessions never set it).
• In session.start(), when _dry_run is set: build the graph (still surfaces node/edge/cycle errors), then raise an internal _StartupProbeComplete to halt cleanly, without requiring the init handshake or starting nodes (no external connections).
• The validator swallows that sentinel and marks the pod ready.

Genuinely broken handlers (node __init__ raising, imports, cycles) still fail validation and keep the pod not-ready, with a real reason. The live session path is untouched.

Result

• Healthy crews: no error, pod correctly reports ready.
• Broken crews: still fail fast with the real cause.
• The docs "ignore this error" note can be removed (follow-up docs PR).

Tests

• Canonical handler (ends with session.start()) passes validation, _ready true.
• Handler whose node __init__ raises fails validation, _ready false with the reason.
• Dry-run start() builds the graph and raises the sentinel with no init event.

Green locally: ruff 0.16.1 (check + format), mypy 1.13.0, pytest (crew suite 26 passed).

Held for review.