Stop the node_fold tests racing the real :4243 — the flake that cost three shard re-runs (#1166) - #1170

…cing the real :4243

`test_refuses_a_live_node_this_process_does_not_own` failed on three unrelated
PRs -- #1162 shard 6, #1164 shard 2, #1169 shard 7 -- each costing an
18-minute shard re-run on a diff it could not see (#1169's was requirements.txt
plus four version strings). CIRISAgent#1166.

Cause: the node port was a bare literal in four places in node_fold, so a test
exercising the ownership logic had to bind the REAL 4243 and then raced
everything else on the box that touches it -- the sibling `own_listener` in
another xdist worker, a leftover node, a TIME_WAIT socket with SO_REUSEADDR.
The flaky-retry plugin re-ran into the same shared environment, so retries
rarely helped.

Worse than flaky: when 4243 was already held, two of the four tests SKIPPED
("port 4243 already in use") -- silently not checked, which is the shape this
very file was written to eliminate.

Now `NODE_FOLD_PORT` names it once; the liveness probe, the ownership check,
the HTTP identity probe and the two operator-facing messages all read it, so
the messages name the port actually probed. An autouse fixture points it at an
ephemeral port the kernel guarantees is unused.

Proven against the failure condition rather than by passing once: with a real
foreign HTTP server squatting on 4243, the old file skipped 2 of 4 tests while
the new one passes all 4 -- serially three times and under `-n 8`. All 44 tests
across the five node_fold-touching files pass; mypy on node_fold is clean (the
two remaining errors are the documented wallet import-untyped ones, identical
on main).

No version bump: #1169 is already taking 2.11.4 and edits the same constants.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01J45bNgkggVC1ntnmp6DqQA

Sonar put this PR's new-code coverage at 71.4% against the 80% gate: of the
lines the port-constant change touched, two sat in `stop_node_fold`'s
"the wheel cannot free the port" arm, which had no test.

That arm is worth one. `stop_node_fold` returns None for BOTH "this process
never started a node" and "this wheel predates shutdown_node", so the warning
is the only thing separating a clean no-op from a node that will still hold
:4243 when the next boot tries to bind it -- the EADDRINUSE the five-platform
gate's post-reset port check exists to catch (CIRISAgent#1102). Both paths are
now asserted, including that the clean one stays quiet.

9 of the 11 changed lines are now covered. The remaining two are inside the
post-boot compose region (node_fold 515-708), which has no test coverage at
all today and is not something this PR introduced.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01J45bNgkggVC1ntnmp6DqQA