Both VirtualTagEngine.Load and ScriptedAlarmEngine.LoadAsync were calling
ScriptEvaluator.Compile directly, bypassing CompiledScriptCache. The
Core.Scripting-008 collectible-ALC fix wired Dispose only through the cache's
Clear()/Dispose(), so the per-publish accretion the -008 fix was meant to
eliminate was still in effect on the actual production path — the headline
'no more restarts needed' guarantee wasn't delivered.
Resolution:
- VirtualTagEngine + ScriptedAlarmEngine each gained a private
CompiledScriptCache<TContext, TResult> instance.
- Both Load methods now call _compileCache.GetOrCompile(source).
- Publish-replace path: _compileCache.Clear() runs alongside the existing
_tags / _alarms clears so the prior generation's ALCs are disposed
before recompile.
- Engine Dispose now calls _compileCache.Dispose() so shutdown actually
releases the emitted assemblies.
Side-fix in CompiledScriptCache: Dispose() set _disposed=true then called
Clear(), but Clear() had a pre-existing 'if (_disposed) return' guard that
aborted the drain unconditionally — making the Dispose-triggered cleanup a
silent no-op. Removed the disposed-guard on Clear() (clearing an empty/
cleared cache is idempotent).
Side-fix in ScriptedAlarmEngine.Dispose: cleared _alarms AFTER the
Task.WhenAll drain. The drain guarantees no background callback is mid-
flight, so clearing is safe. Previously _alarms was deliberately NOT
cleared on Dispose (per Core.ScriptedAlarms-005), but that left the
AlarmState records holding TimedScriptEvaluator → ScriptEvaluator → delegate
references that rooted the emitted assemblies, defeating the cache's
Dispose work on the engine side.
Regression tests:
- VirtualTagEngineTests.Dispose_unloads_compiled_script_assembly
- ScriptedAlarmEngineTests.Dispose_unloads_compiled_predicate_assembly
Both use WeakReference + bounded GC.Collect() to prove the emitted
assembly is reclaimable after engine.Dispose(). The alarms test had to
be synchronous (not 'async Task<WeakReference>') because async state
machines capture locals as state-struct fields, keeping them alive past
the method's apparent end and defeating GC.
Verification:
- Core.Scripting.Tests: 104/104 (unchanged).
- VirtualTags.Tests: 57/57 (was 56 — +1 unload test).
- ScriptedAlarms.Tests: 67/67 (was 66 — +1 unload test).
- All other consumer suites still green.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>