feat(preflight): replace SUPPORTED_RUNTIMES static list with adapter discovery

[HongmingWang-Rabbit]

Summary

Closes task #123 — last piece of #87 cleanup independent of the cli_executor.py template moves.

Pre-fix: workspace/preflight.py:11 hardcoded a tuple of "supported" runtime names. Every new template repo required a code change in molecule-runtime to be recognized — direct violation of the universal-runtime principle (#87).

Post-fix: discovery-based validation via the same `ADAPTER_MODULE` env var that production load paths already use (`workspace/adapters/init.py:get_adapter`). Distinguished failure modes so operator messages are concrete:

Mode	Detail
`ADAPTER_MODULE` unset	"no adapter installed; set the env var"
Module won't import	Underlying ImportError type + message
No `Adapter` class	"Add Adapter = YourClass per BaseAdapter convention"
`Adapter.name()` raises	"name() raised : " — must be side-effect-free
`Adapter.name()` returns non-string	"must return a non-empty string"
`Adapter.name() != config.runtime`	WARNING (not fail) — adapter wins, config drift surfaced

Behavioral change worth calling out

The drift case is a softening: prior static-list path hard-failed config.runtime values not in the allowlist; now drift is a warning so the workspace still boots while operator gets actionable info naming both the configured + installed runtime names. Reflects production reality (the adapter wins regardless).

Tests (6 new + autouse fixture)

• Replaces obsolete "static list pass/fail" tests with 6 cases covering each distinguished failure mode + a positive happy-path test
• Autouse `_default_langgraph_adapter` fixture pre-installs a fake adapter via `sys.modules` monkey-patching, so existing tests building default WorkspaceConfig (`runtime="langgraph"`) inherit a valid adapter
• Failure-mode tests opt out via `@pytest.mark.no_default_adapter` (registered in pytest.ini)
• Sentinel pattern (`_UNSET = object()`) for `name_returns` so None / int / etc. are passable test values

Verification

• 22/22 preflight tests pass (was 16; +6 new failure-path tests)
• 1256/1256 workspace pytest pass (was 1251; +5 net)
• No production code path other than preflight changed

🤖 Generated with Claude Code

[HongmingWang-Rabbit]

Code review (5-axis) — needs cross-account approval

Note: posting as a comment because gh's active account is also the PR author. The user's other account will need to formally approve.

Correctness: ✓

• Each failure mode (unset / unimportable / no Adapter class / name() raises / non-string name() / drift) gets a distinguished operator-facing message — debugging is concrete, not a single generic "unsupported"
• Drift case correctly returns (True, detail) so the dispatch in run_preflight puts it in warnings instead of failures
• Empty-string detail on success skips the warning branch — clean signal/no-noise

Tests: ✓

• 6 new failure-path tests, each pinned to the exact substring of the operator message — a future refactor that drops the actionable detail will fail loudly
• _UNSET = object() sentinel lets None be a passable test value for name_returns — exactly the bug a if name_returns is not None check would have caused
• Autouse fake-adapter fixture + @pytest.mark.no_default_adapter opt-out keeps existing tests untouched while exercising failure paths cleanly

Architecture: ✓

• Aligns with chore(template): add 4 evolution crons — ecosystem / plugins / template / channels #87 universal-runtime principle (cited in body) — adapter is now source of truth, config.yaml is just a label
• Reuses the same ADAPTER_MODULE env convention as workspace/adapters/__init__.py:get_adapter — consistent pattern, no new contract

Security: ✓

• importlib.import_module(env_var) is arbitrary-module-loading by design, but ADAPTER_MODULE is image-build-time config (Dockerfile ENV), not user input
• Same trust boundary as the existing production load path

Performance: ✓ — preflight runs once at startup; one module import is negligible

---

Nit: install_fake_adapter uses id(monkeypatch) in the module name to avoid sys.modules collisions across tests, but pytest's monkeypatch already restores sys.modules at teardown. The unique-name suffix is defensive overkill — could just be f"_fake_adapter_{name.replace('-', '_')}". Not worth blocking on; current code is correct, just slightly more machinery than needed.

Verdict: would approve. Auto-merge already armed; will fire once CI greens + an approving review lands.