refactor(memory): delete dead sidecar infrastructure and HardScope plumbing

[Aaronontheweb]

Summary

Part one of a multi-step cleanup of the Netclaw memory subsystem. This PR deletes dead code only — no behavior changes for any test that was already passing, no schema migrations, no API surface that users could reach. It's a pure net-removal refactor that leaves the memory system easier to reason about before checkpoint 2 (the Domain field + schema migration) lands.

Background

While closing #582 I discovered that the memory subsystem had two large pieces of dead infrastructure sitting as "scheduled for removal" tombstones since late March:

1. Recall / observation sidecars. Introduced as a prototype in PR feat(memory): add observer-driven recall planning and retrieval prototype #198 on 2026-03-12, replaced by DeterministicCandidateSelector for recall and by SessionMemoryObserverActor.RunDistillationAsync for memory formation within ~2 weeks. The infrastructure got an explicit "Scheduled for removal" annotation on 2026-03-25 (PR refactor: consolidate SessionConfig — reduce 16 properties to ~5 meaningful settings #414) but the deletion never shipped. Zero production code paths reach any of it.
2. HardScope / HardScopeOverride / domain-derived boundary inference. DeterministicRetrievalRequestPlanner.ResolveHardScope parsed a session ID prefix into a per-project domain, but the coordinator always populated HardScopeOverride first via Protocol.SessionId.ToMemoryDomain(), which unconditionally returns SecurityPolicyDefaults.DefaultMemoryDomain. The session-ID parsing branch was unreachable. Domain affinity in DeterministicCandidateSelector was disabled in fix(memory): add composite-score floor to recall ranker (#582) #585 for the same reason.

Beyond those two, the code review pass also flagged a handful of small residuals: dead parameters on MemoryPolicyScopeResolver.ResolveBoundary, a duplicate private helper in the coordinator, dead message records in LlmMessages, and two WHERE 1 = 1 SQL placeholders. Those are cleaned up in the second commit.

Design decision: "audience + lexical/facet/anchor matching, nothing else"

Before touching anything I verified with the project owner that the mental model going forward is:

• Audience (Public ⊂ Team ⊂ Personal) is the only security / isolation axis. MemoryPolicyEvaluator.AllowedAudienceWireValues already does the correct walled-garden check, including the case where an operator in a privileged DM context should be able to reference memories formed in a team channel.
• Lexical + facet + anchor matching is the only relevance signal. No hard scope, no channel mapping, no per-project taxonomy. Issue Memory domain scoping is half-implemented: all sessions collapse to project:default #584 captures why per-project taxonomy is a bad fit for this system: channel-based hard scoping would actively break operator cross-channel workflows that the audience hierarchy already supports correctly.
• Composite-score floor from Memory recall unconditionally injects top-N items with no score floor, polluting unrelated sessions #582 stays as the noise gate.

That framing is the basis for every deletion in this PR. If a concept only existed to prop up one of the three things above (hard scope, LLM-driven sidecars, domain-derived boundaries) it's gone.

What this PR does

Sidecar infrastructure deleted

File	Disposition
src/Netclaw.Actors/Sessions/SessionSidecarRunner.cs	Deleted entire file (321 lines) — LLM sidecar call executor, only reachable from the dead recall fallback path
src/Netclaw.Actors/Sessions/MemorySidecarPromptBuilder.cs	Reduced from 151 lines to the single BuildClassificationRules helper still used by the live SessionMemoryObserverActor distillation path
src/Netclaw.Actors/Sessions/SidecarRecallPlanner.cs	SidecarMemoryObserver class deleted (observation sidecar). SidecarRecallPlanner class retained — still used by the live SqliteFindMemoriesTool to build structural request plans without any LLM involvement. domain parameter dropped from BuildRequest signature.
src/Netclaw.Actors/Sessions/MemorySidecarContracts.cs	MemoryObservationRequest, MemoryObservationCurrentTurn, MemoryObservationRecentContext, MemoryObservationPolicyScope, MemoryObservationCompleted, RecallPlanningCompleted all deleted. MemoryProposal, MemoryAnchor, MemoryRelation, MemoryProposalOperation, RecallPlanningRequest, RecallQueryPlan retained — all live via the distillation path or the find_memories tool.
src/Netclaw.Actors/Sessions/SQLiteMemoryRecallCoordinator.cs	The entire if (!_sessionTuning.MemorySidecarsEnabled) … fallback branch in RecallAsync, plus BuildPlanAsync, LogRecallTrace, TokenizeTerms, FallbackSearchTerms, NormalizeRequest. Constructor parameters reduced from 6 (store, logger, clientProvider, sidecarPlanner, recallPlanGate, sessionTuning, sessionConfig) to 3 (store, logger, sessionTuning).
src/Netclaw.Actors/Sessions/LlmSessionActor.cs	Command<MemoryObservationCompleted> handler deleted — registered handler with zero senders.
src/Netclaw.Actors/Sessions/LlmMessages.cs	MemoryObservationFailed and RecallPlanningFailed records deleted — also had no senders.
src/Netclaw.Configuration/SessionTuning.cs	MemorySidecarsEnabled flag deleted. Config binding in SessionConfig, JSON schema entry in netclaw-config.v1.schema.json, and the Memory_sidecars_enabled_by_default default-check test in SessionConfigDefaultsTests all removed to match.

HardScope / domain plumbing deleted

File	Disposition
src/Netclaw.Actors/Sessions/DeterministicRetrievalPlanning.cs	ResolveHardScope and IsTransportScopedSession deleted. HardScope property removed from DeterministicRetrievalRequestPlan.
src/Netclaw.Actors/Sessions/IMemoryRecallCoordinator.cs	HardScopeOverride field removed from AutomaticRecallRequest.
src/Netclaw.Actors/Sessions/Pipelines/SessionRecallManager.cs	Dropped the HardScopeOverride: sessionId.ToMemoryDomain() argument from the recall request construction.

Boundary resolution simplified

All Netclaw memory runs inside a single trusted-instance daemon. InferLegacyBoundaryFromDomain ultimately returned TrustedInstanceBoundary for every input the system actually produces. Rather than thread that fiction through 7 call sites:

• Every call to SecurityPolicyDefaults.InferLegacyBoundaryFromDomain(...) in the recall path replaced with SecurityPolicyDefaults.TrustedInstanceBoundary directly. Curation-pipeline call sites are still on the old function and are deferred to checkpoint 2.
• MemoryPolicyScopeResolver.ResolveBoundary simplified to the single-value case: if a caller-supplied boundary is present, trim and return it; otherwise return TrustedInstanceBoundary. The audience and sessionId parameters became dead weight and were dropped from the signature and every call site (curation pipeline, policy gates, all four SQLite memory tools).
• The private SQLiteMemoryRecallCoordinator.ResolveBoundary helper was byte-identical to the scope-resolver method after simplification — inlined at the call site and deleted.

Search method collapse

SQLiteMemoryStore.SearchByPlanAsync (the domain-scoped variant) and SearchAcrossDomainsByPlanAsync were merged into the single cross-domain method. The domain parameter is gone, the WHERE d.domain = $domain / WHERE r.domain = $domain SQL clauses are gone, and two WHERE 1 = 1 scaffolding artifacts left behind by the inlining were also removed.

Test fallout

• Deleted entire files: MemorySidecarPromptBuilderTests.cs (tested the deleted prompt builder methods) and SessionSidecarRunnerTests.cs (tested the deleted runner).
• Deleted specific tests: Sidecar_observation_promotes_strong_user_assertion_into_memory from LlmSessionIntegrationTests, plus the "memory observation sidecar" branch of the fake chat client that only existed to serve that test.
• Parameter cleanup in ~17 test files: MemorySidecarsEnabled = false/true literals stripped from every SessionTuning constructor call. HardScopeOverride: "user:aaron" dropped from DeterministicRetrievalPlanningTests, MemoryRedesignedEvalSuiteTests. plan.HardScope assertions removed.
• DeterministicCandidateSelectorTests fixture update: MakePlan lost its hardScope parameter. The Score_geometry_domain_affinity_boosts_same_domain_match test from fix(memory): add composite-score floor to recall ranker (#582) #585 was replaced with Domain_is_not_a_scoring_signal which asserts same-domain and cross-domain candidates score identically.
• SQL call site updates: SQLiteMemoryStoreTests and MemoryRedesignedEvalSuiteTests switched from SearchByPlanAsync(..., "project:test", ..., InferLegacyBoundaryFromDomain(...), ...) to SearchAcrossDomainsByPlanAsync(..., TrustedInstanceBoundary, ...).

What this PR deliberately does NOT do

These are all tracked as checkpoint 2 work:

• Protocol.SessionId.ToMemoryDomain() is still a trivial constant passthrough with ~14 call sites. Inlining it is mechanical churn across LlmSessionActor, SessionMemoryObserverActor, the SQLite memory tools, SubAgentActor, and SessionToolExecutionPipeline.
• The Domain field on SQLiteMemoryDocument, SQLiteMemoryHydratedItem, SQLiteMemorySearchResult, SQLiteMemoryCurationOperation, SQLiteMemoryAnchor, AutomaticRecallItem, and MemoryCheckpointPayload stays for now. Removing it touches every writer and reader of every memory record type.
• The domain column on memory_documents, memory_records, memory_anchors, memory_edges stays. Dropping it needs an idempotent schema migration path, and schema migration risk is better decoupled from this deletion-heavy PR.
• NormalizeLegacyBoundariesAsync still reads the domain column to bucket pre-existing data. Dead once the column is gone.
• SecurityPolicyDefaults.InferLegacyBoundaryFromDomain still has two callers in MemoryCurationPipeline (lines 319, 508) that haven't been migrated yet. SecurityPolicyDefaults.DefaultMemoryDomain stays as the constant the column receives.
• openspec/specs/netclaw-agent-memory/spec.md still references domain as part of the memory policy envelope and has a scenario about domain=business filtering. The OpenSpec update will land alongside the checkpoint 2 PR so the spec describes the final landed state in one pass.

Net effect

~1,450 lines deleted, ~80 added across two commits. All 1,967 tests across 7 test projects pass locally. Slopwatch clean (0 issues). dotnet build clean (0 warnings).

Closes #584 in spirit — the decision captured there is "domain scoping is the wrong mental model; audience handles it." The formal issue close will follow the checkpoint 2 PR that finishes removing the vestigial column and field.

Test plan

• dotnet build green across all projects (0 warnings, 0 errors)
• dotnet test src/Netclaw.Actors.Tests/Netclaw.Actors.Tests.csproj — 840 passed / 0 failed
• dotnet test full solution — 1,967 passed / 0 failed across 7 test projects
• dotnet slopwatch analyze — 0 issues
• CI green on this PR