feat(l3): dedicated l3Llm config slot for abstraction pass

[chiefmojo]

Problem

L3 abstraction runs on the main llm. On cheap models (e.g. gemini-2.5-flash-lite) this over-extracts and truncates the JSON output, producing 'constraints'/'inference' must be an array errors. The slot is quality-sensitive and runs off the turn-response path — a slower but more capable model here improves world-model quality with no impact on latency.

Changes

Adds an l3Llm top-level config slot using the same SkillEvolverSchema shape as skillEvolver. When blank (the default), behavior is identical to today — deps.l3Llm ?? deps.llm falls through to the main LLM. When set, the L3 clustering → abstraction subscriber uses the dedicated client.

l3Llm:
  provider: openai_compatible
  endpoint: https://openrouter.ai/api/v1
  model: anthropic/claude-sonnet-4-5
  apiKey: sk-or-...

• defaults.ts — blank default (inherits main llm)
• schema.ts — l3Llm: SkillEvolverSchema entry
• memory-core.ts — bootstrap client (mirrors reflectLlm path; errors log, don't crash)
• deps.ts / types.ts — PipelineDeps.l3Llm and PipelineHandle.l3Llm fields
• orchestrator.ts — threads the field through to the handle
• orchestrator.test.ts — two tests: dedicated client threads through, null when unconfigured

Config example

Operators running a budget model as their main llm can now point L3 at a stronger model:

llm:
  provider: openai_compatible
  model: gemini-2.5-flash-lite  # cheap, fast, good for scoring

l3Llm:
  provider: openai_compatible
  model: anthropic/claude-sonnet-4-5  # slow ok — L3 is async
  apiKey: sk-or-...

[itxaiohanglover]

Good config design! Dedicated l3Llm config slot allows using a stronger model for L3 abstraction without affecting companion latency. The fallback to shared llm settings is sensible. Adding l3Llm.apiKey to SECRET_FIELD_PATHS is correct.

[Memtensor-AI]

Automated Test Results: PASSED

Cloud test-engine rerun against dev-v2.0.22 completed successfully.

• Run: tr-7220330f-3c8 on cloud test-engine 10010
• memos_local_plugin/unit: 39 passed, 0 failed, 0 skipped

Manual code review is still required before merge.

[Memtensor-AI]

✅ Automated Test Results: PASSED

All tests passed (43/43 executed). memos_local_plugin/unit: 43/43. Duration: 12s [advisory, non-gating] AI-generated tests on branch test/auto-gen-05768d70954acbe1-20260709111339: 51/51 passed — these do NOT affect the PR verdict; review the branch manually.

Branch: feat/l3-dedicated-llm

[Memtensor-AI]

✅ Automated Test Results: PASSED

All tests passed (43/43 executed). memos_local_plugin/unit: 43/43. Duration: 12s [advisory, non-gating] AI-generated tests on branch test/auto-gen-41b2127f5d64fbe6-20260709120727: 77/77 passed — these do NOT affect the PR verdict; review the branch manually.

Branch: feat/l3-dedicated-llm