Fix peer.py catching httpx exceptions around requests calls (#131)

[ClaydeCode]

Fixes #131.

Problem

shard_core/service/peer.py update_peer_meta calls requests.get(...) and then response.raise_for_status(), which raises requests.exceptions.HTTPError — but the handler caught httpx.HTTPStatusError. So any non-2xx peer response escaped the handler, propagated through the asyncio.gather in update_all_peer_pubkeys (which had no return_exceptions), and aborted the whole peer-refresh cycle for that tick (only contained by PeriodicTask retrying a minute later).

Change

• Catch requests.HTTPError instead of httpx.HTTPStatusError, so a non-2xx peer response marks that peer unreachable (as intended) instead of escaping. Dropped the now-unused httpx import.
• Add return_exceptions=True to the gather, with a per-peer warning log, so any unexpected exception from one peer (e.g. a wrong-identity KeyError) can no longer abort refreshing the others.
• New tests/test_peer_refresh.py with two regression tests over update_all_peer_pubkeys.

Verification

tests/test_peer_refresh.py — 2 passed. Full suite: 227 passed; the one unrelated failure (test_app_lifecycle.py::test_app_starts_and_stops, a Docker container-status timing test) also fails on a clean origin/main on this runner, so it is pre-existing and not caused by this change.

Both new tests were proven to have teeth (each fails when the corresponding half of the fix is reverted):

• revert the exception type → test_unreachable_peer_does_not_block_others fails (bad peer no longer marked unreachable),
• remove return_exceptions=True → test_unexpected_peer_error_does_not_block_others fails (gather re-raises),
• remove the warning log → test_unexpected_peer_error_does_not_block_others fails (warning assertion).

Review panel

Three reviewers ran (adversarial always-on; test-adversary because the diff adds real logic + tests; DevEx because it touches tests). No blocking findings from any. Advisories addressed:

• Decorative is_reachable is True assertion (test-adversary, adversarial, DevEx) — fixed: the good peer is now inserted is_reachable=False, so the success path's flip to True is actually asserted.
• Test 1 couldn't distinguish the HTTPError path from a ConnectionError fallback (test-adversary) — fixed: assert len(rsps.calls) == 2 pins that both whoareyou requests fired.
• Test 2 didn't assert isolation to the broken peer, and the new warning log was unasserted (test-adversary, DevEx) — fixed: assert the broken peer is left untouched and that a warning containing its short_id was emitted.
• Exception-alias style inconsistency (DevEx) — fixed: use requests.HTTPError to read symmetrically with the adjacent requests.ConnectionError handler.

Advisories intentionally not addressed (out of scope for #131, pre-existing): other requests failure modes (Timeout, SSLError) don't mark a peer unreachable; _on_peer_write calls update_peer_meta outside the guarded gather. Neither is introduced or worsened by this diff.

🤖 Generated with Claude Code