sheawinkler/ContextLattice

ContextLattice v3.3.4

Target commit: a8cb9c3 (public/main)

What changed since v3.3.3

• Synced calibrated resource requirements and lane-sizing guidance into public docs/README.
• Added explicit operator guidance for keeping RAM/storage in safe ranges while preserving full feature behavior.
• No API contract regressions; this is an operational/docs refresh release.

Low-Usage Settings Playbook (no feature degradation)

Use these settings to stay near the low end of each lane range without disabling capabilities.

| Setting | Recommended value | Why it lowers usage | Feature impact |
| --- | --- | --- | --- |
| COMPOSE_PROFILES | core,llm,observability (default) | Avoids loading spike-lab adapters unless needed | None (full core product remains active) |
| Spike-lab activation | Enable only during benchmarks/spikes | Prevents idle adapter memory/CPU overhead | None for normal recall/runtime |
| ORCH_RETRIEVAL_MODE_DEFAULT | balanced | Keeps source fanout/limits moderate by default | None (deep still available per request) |
| ORCH_RETRIEVAL_ENABLE_STAGED_FETCH | true | Returns fast lanes first; slow lanes continue in background | None (slow lanes still contribute via continuation) |
| ORCH_RETRIEVAL_FAIL_OPEN_TIMEOUT_CONTINUATION_ENABLED | true | Prevents blocking on slow lanes while still warming/cache-populating | None |
| GO_RETRIEVAL_ADAPTIVE_TIMEOUT_ENABLED + ORCH_RETRIEVAL_ADAPTIVE_TIMEOUT_POLICY_ENABLED | true | Auto-trims timeout pressure under load | None |
| GO_RETRIEVAL_CONTINUATION_MAX_INFLIGHT | 8 | Caps background continuation concurrency | None (continuations still run) |
| GO_RETRIEVAL_CONTINUATION_MAX_INFLIGHT_PER_SOURCE* | 2 | Prevents one source from monopolizing memory/threads | None |
| ORCH_STORAGE_GOVERNANCE_MIN_FREE_GB | 40 (or higher) | Maintains free-space safety margin | None |
| Telemetry retention/compression (GO_TELEMETRY_RETENTION_DAYS=75, blob compression+GC enabled) | Keep defaults | Reduces footprint growth over time | None (non-telemetry learning retained) |

Lane resource ranges (current public guidance)

| Runtime lane | CPU | RAM | Storage |
| --- | --- | --- | --- |
| HF/Glama lite (single container) | 2-4 vCPU | 4-8 GB | 20-50 GB SSD |
| Local Lite compose | 2-4 vCPU | 8-12 GB | 25-80 GB SSD |
| Local Full (no spike-lab) | 6-8 vCPU | 12-20 GB | 100-180 GB SSD |
| Local Full + spike-lab | 8-12 vCPU | 24-32 GB | 180-300 GB SSD/NVMe |

Operational note

• Recent calibrated sample: Full + spike-lab observed ~16.39 GiB container RSS; Full baseline (no spike-lab adapters) observed ~7.70 GiB.