fix(memory-core): probe local LLM capabilities

[gavinlouuu-kpt]

Summary

• add explicit ollama and llama.cpp standalone backend profiles with native startup probes
• expose safe configured/effective context limits and model readiness in /health
• refuse oversized prompts before a compatible backend can silently truncate them
• support explicit reasoning/Jinja and provider-specific request body controls without runner edits
• preserve all new controls across gateway, memory, skill-extractor, and offload call paths

Verification

• npm test (5 tests: Ollama and llama.cpp HTTP integrations, failed-model redaction, prompt budget, YAML pass-through)
• npm run build:plugin
• git diff --check

Tracks downstream CDB issue gavinlouuu-kpt/codex-discord-bridge#248.

[gavinlouuu-kpt]

Downstream staging proof for CDB #248:

• npm test: 6/6 passed (Ollama + llama.cpp HTTP integrations)
• plugin build and Docker rebuild passed
• strict built-image llama.cpp canary: /health.status=ok, model ready, effective context 8192, effective input budget 7168
• staging deployment is Docker-healthy with restart count 0 and now safely reports its pre-existing missing backend/config incompatibility as degraded
• production container/image was not touched

Evidence SHA-256: 3d88aab594ba62e18ee54f9be14cbfeff515a5a2654e233f81d03cdd89b38d86.

[gavinlouuu-kpt]

Live downstream staging validation passed at head ef95f3f. Ollama 0.30.7 loaded qwen3.6:27b; health now reports configured/effective context 8192 and input budget 6144 instead of the probed 262144 training limit. Real L0-to-L1 extraction and cross-session recall succeeded. Evidence SHA-256 f4d2e1de8fe430e3ec10fac87980abd0469339979fc5fa7c19bb37a0577f1215.

[Maxwell-Code07]

Thank you so much for your attention and contribution! We will arrange an internal review for this PR shortly, and all feedback will be shared right here in the discussion.