Three review findings + a real bug:

1. Replace 'Effect.promise(() => Flock.withLock(... runPromise(fetchApi) ...))'
   with 'Effect.scoped(Effect.gen { Flock.effect(lockKey); fetchAndWrite })'.
   Flock.effect is the existing scoped acquire/release; the runPromise
   re-entry inside the Promise lock body was dropping the parent fiber's
   tracing span, scope, and interruption.

2. Replace hand-rolled 'ManagedRuntime.make(defaultLayer, { memoMap })' with
   the existing 'makeRuntime(Service, defaultLayer)' helper from
   src/effect/run-service.ts. Same pattern as bus, sync, instance-store, etc.

3. Use 'Effect.cachedInvalidateWithTTL(populate, Duration.infinity)' for
   single-flight cache + manual invalidation. Concurrent first-callers now
   share one populate; refresh() invalidates so the next get() re-populates.

4. Restore eager initial refresh — Schedule.fixed waits the duration before
   the first run, which would have lost the original 'void refresh()' on
   module load. Now: refresh().pipe(Effect.andThen(refresh().repeat(...))).

5. Simplify fresh() mtime check — fs.stat's info.mtime is always Option<Date>;
   dropped the dead Option.isOption guard.

6. Drop unnecessary Effect.fnUntraced on loadFromDisk/loadSnapshot (these
   are values, not traced effects). Replaced loadSnapshot's manual try/catch
   wrapper with Effect.tryPromise + Effect.catch.

7. Effect.orDie on populate — Service interface declares get returns
   Effect<Record<string, Provider>> (E = never); failures from HttpClient,
   FileSystem, JSON.parse must die rather than silently leak.

Smoke + provider tests pass.

- Use Schedule.spaced (not fixed) so the periodic refresh waits between
  completions instead of firing twice on startup.
- Pull JSON.parse out of the Flock-held scope — keeps the cross-process
  lock release-eager.
- Only invalidate after a successful fetchAndWrite; the no-op fresh-skip
  and ignored-failure paths shouldn't force the next get() to re-read.
- Drop comments narrating obvious behavior; fix a misplaced WHY comment.
- Use HttpClient.make (matches mockHttpClient elsewhere) instead of
  makeWith with a Preprocess cast in the new tests.

Mutating Flag.OPENCODE_MODELS_PATH at module-load time leaked the change
into every subsequent test file in the same bun process — Provider state
init then read undefined, fell into the snapshot/fetch fallback, and on
slower CI lanes timed out or returned empty providers (surfacing as the
"no providers found" failure on Windows for the SDK no-reply parity test).

Wrap both mutations in beforeAll/afterAll so they only apply while this
suite runs.