feat(smoke): consult runtime_wedge after execute() to catch SDK init wedges

[HongmingWang-Rabbit]

Summary

• Adds a runtime_wedge.is_wedged() consultation at the end of every result path in run_executor_smoke
• Upgrades the provisional PASS-on-timeout / PASS-on-non-import-error branches to FAIL when an adapter has marked the runtime wedged
• Closes the gap that let PR-25-class regressions (claude-agent-sdk init wedge from a malformed CLI argv) ship to GHCR — the SDK's 60s initialize() timeout used to read as "network boundary, imports healthy" through the 5s outer wait_for

Why

Today's incident: a wheel-side change (template repo PR #25) shipped extra_args={"dangerously-load-development-channels": None}. The CLI rejected the bare --flag, the SDK timed out at initialize, every A2A turn wedged. Workspace dd40faf8 dead-locked.

The existing import-smoke (issue #2275, task #126) catches lazy-import regressions but treats every other timeout / non-ImportError as PASS. That was an explicit choice — adapter-level errors are caught by adapter-level tests — but it makes the publish gate blind to wedges that surface as the executor returning a 'success-shaped' control flow.

Design

Universal, opt-in via the existing runtime_wedge module (already documented as the cross-cutting wedge state holder; already read by heartbeat to flip the workspace to degraded). Adapters call runtime_wedge.mark_wedged(reason) from their executor's wedge catch arm; the smoke reads is_wedged() and overrides PASS → FAIL.

• Additive: adapters that don't integrate with runtime_wedge keep working as before.
• Fail-open on import: a corrupt install where runtime_wedge is missing reads as 'no wedge info' (same posture heartbeat takes), so observability never crashes the smoke gate.
• Calibrated timeout: CI usage example sets MOLECULE_SMOKE_TIMEOUT_SECS=90 to outlast the SDK's 60s initialize() handshake — comments call this out so a future contributor doesn't shrink it accidentally.

Test plan

• pytest tests/test_smoke_mode.py — 28 passed, 2 skipped (real-SDK-only)
• pytest tests/test_smoke_mode.py tests/test_runtime_wedge.py tests/test_heartbeat.py — 57 passed, 2 skipped
• Regression-injection: silencing wedge_reason_str = _check_runtime_wedge() makes both wedge-shape tests fail at unit-test time
• Wired into publish-template-image.yml smoke step in a follow-up molecule-ci PR (this PR ships the runtime-side change so the wheel publishes first)
• Wired into claude-code template's _mark_sdk_wedged in a follow-up template-repo PR (so the heartbeat + this gate both see the same wedge)

Follow-ups (separate PRs by design — this PR is the foundation):

1. molecule-ci: extend publish-template-image.yml to run a MOLECULE_SMOKE_MODE=1 boot with MOLECULE_SMOKE_TIMEOUT_SECS=90 between import-smoke and GHCR push
2. molecule-ai-workspace-template-claude-code: have _mark_sdk_wedged also call runtime_wedge.mark_wedged so the universal hook sees claude wedges (heartbeat already needs this — the local sticky flag isn't currently observed)