fix(ci): falcon-hello-demo raced on a hardcoded port — the flake that failed #365

[avrabe]

Verify-Filter: (has-tag "v0.1")

#365's gate failed on one step out of 163, and not because of anything in #365:

[ FAIL (rc=1)] FV-FALCON-WORLD-001: scripts/falcon-hello-demo.sh
  [falcon-hello-demo] vehicle sent 0 heartbeat(s)
  FAIL: expected at least 8 heartbeats, got 0
  error: bind 127.0.0.1:14701: Address already in use (os error 98)

The demo hardcoded the 14700/14701 pair, the gate runners are shared, and two
jobs on one runner fight for the same socket.

This is the worst shape a flake can take on a required check: it fails loudly, names
a plausible functional symptom ("got 0 heartbeats"), and the real cause is one line
further down. #365 promotes 105 artifacts, so the natural reading was that a
promotion had broken something.

Fix

Probe for a free pair instead of assuming one. FALCON_HELLO_PORT_BASE still wins
so a bench operator can pin it; if 64 probes all fail it falls back to the historical
default rather than erroring — a demo that can't find a port should still try, not
add a second failure mode.

Measured, both directions, 3 concurrent runs

	result
forced to the old fixed base	2 of 3 FAIL — Address already in use ×3, got 0 heartbeats ×2
with the probe	3 of 3 PASS, 14 heartbeats each

The negative case is #365's exact error text reproduced from the exact condition,
not a synthetic stand-in.

Refs #365, #345

🤖 Generated with Claude Code

https://claude.ai/code/session_01HvusAXYbHLyv3uTzfBcMbG