v0.4.5 — Apache-2.0 relicense, fossic PyPI

v0.4.5 — Apache-2.0 relicense and fossic PyPI switch

Ecosystem hygiene release. No behavior changes from v0.4.4.

Highlights

License

• Relicensed to Apache-2.0 (was MIT). See LICENSE.
• SPDX identifiers added to all source files.
• CONTRIBUTING.md now requires DCO signoff (git commit -s).

Fossic dependency

• Fossic now installs from PyPI as an optional extra:
  pip install cerebra → SQLite-only mode
  pip install cerebra[fossic] → adds run-cycle, serve, event streams

• requires-python widened to >=3.12,<3.14 (fossic 1.8.3 ships cp313 wheels).

Repo hygiene

• Documentation restructure: HISTORY.md, ARCHITECTURE_STATE.md, TECH_DEBT.md replace pre-cleanup archive framing.
• Internal-artifact .gitignore patterns added.
• CI installs with [dev,fossic] extras.

Tests

1686 tests passing. Coverage: 82.30%.

Migration

None required. Users of prior versions continue to work as before.

New users install via:

uv sync --extra dev --extra fossic

Cerebra pre-dyson-sphere v0.4.4

Cerebra pre-dyson-sphere baseline — v0.4.4

The archival snapshot of Cerebra at
2026-06-21, captured immediately after the
fossic v1.6.0 substrate release and
immediately before the dyson-sphere migration absorbs portions of the
persistence layer into a Rust-backed event-sourced substrate.

This release exists so the before state remains inspectable,
runnable, and citable for architectural comparison and academic
reference.

What's shipped at v0.4.4

A local-first cognitive cycle runtime built in Python — not a RAG pipeline,
not a chatbot wrapper. Each cycle retrieves context from an ingested
knowledge vault, calls a local LLM (Ollama), evaluates the output across six
epistemic signals, routes the next action through a rule engine (Clutch),
optionally escalates to a bandit-driven strategy selector (Catalyst), writes
the result as a dual-format episode, and decides whether to continue,
recurse, or stop. Every action leaves an inspectable trace.

Runtime surface:

• 21-command CLI (init, ingest, search, context, run-cycle,
  classify, inspect, export, lifecycle, serve, …)
• Hybrid retrieval — lexical (FTS5) + vector (mxbai-embed-large-v1) +
  SKU-shaped + graph-expanded; composite-scored
• ClutchEngine — priority-rule controller with hysteresis, mode persistence,
  cascade depth
• CatalystEngine — bandit-driven arm selection (epsilon-greedy + UCB1) over
  five cognitive strategy arms
• TruthTower — five-tier derived workspace (T1 Evidence → T5 Goal)
• Re-injection loop — continuation bundles for cognitive continuity across
  context-window limits
• Six epistemic signal evaluators — coherence, groundedness, relevance,
  precision, generativity, epistemic humility
• Leeway network — three-tier safety architecture (constitutional,
  capability, conditional grants)
• Inspector CLI — forensic query surface over events, sessions, cycles,
  signals, leeway
• Dual persistence — SQLite (18 schema migrations) + FossicStore
  causation-chained event streams
• Graph export to visualization-compatible JSON

Build & test discipline

• 14 sequential build phases (v0.0 → v0.4.4), each verified by tests
  before proceeding
• ~19,000 LOC of Python in cerebra/; ~29,000 LOC of tests
• mypy in --strict mode, only 4 # type: ignore suppressions in the
  whole package
• Coverage gate at 80% (--cov-fail-under=80)
• Tool versions pinned exactly to match the dev environment:
  ruff==0.4.8, black==24.4.2, mypy==1.10.0

Setup

git clone https://github.com/bitmosh/cerebra-classic
cd cerebra-classic
pip install -e ".[dev]"

Prerequisites: Python 3.12+, Ollama running locally with a model pulled
(e.g. ollama pull granite3.2:3b). The fossic Rust extension ships as a
pre-built wheel under vendor/ — no Rust toolchain needed for the supported
platform (CPython 3.12 on Linux x86_64 with glibc ≥ 2.34). See
[vendor/README.md](https://github.com/bitmosh/cerebra-classic/blob/claude/github-docs-cleanup-hc15zm/vendor/README.md) for rebuild instructions.

Quick demo:

cerebra init examples/demo-vault --force
cerebra ingest examples/docs --vault examples/demo-vault
cerebra run-cycle simple.planning.v0 \
  --goal "What are the key patterns in knowledge-intensive systems?" \
  --vault examples/demo-vault
cerebra inspect cycle show --signals --vault examples/demo-vault
Maintenance policy
This fork accepts critical security patches, documentation corrections, and
dependency pin updates. It rejects new features, architectural changes,
performance improvements, and cosmetic refactors — those land in the live
[Cerebra](https://github.com/bitmosh/cerebra) repository.

Known issues
[#2](https://github.com/bitmosh/cerebra-classic/issues/2) — dedup_siblings
SKU-match routing is dead code due to a split-separator mismatch in
[cerebra/retrieval/lattice_dedup.py](https://github.com/bitmosh/cerebra-classic/blob/claude/github-docs-cleanup-hc15zm/cerebra/retrieval/lattice_dedup.py). Function still returns a valid
winner via composite-score routing; SKU-aware routing has never fired.
Full list: [docs/KNOWN_ISSUES.md](https://github.com/bitmosh/cerebra-classic/blob/claude/github-docs-cleanup-hc15zm/docs/KNOWN_ISSUES.md).

Documentation
[README.md](https://github.com/bitmosh/cerebra-classic/blob/claude/github-docs-cleanup-hc15zm/README.md) — entry point, setup, demo
[docs/CEREBRA_CLASSIC.md](https://github.com/bitmosh/cerebra-classic/blob/claude/github-docs-cleanup-hc15zm/docs/CEREBRA_CLASSIC.md) — development arc,
current state, why archived as classic
[docs/ARCHITECTURE.md](https://github.com/bitmosh/cerebra-classic/blob/claude/github-docs-cleanup-hc15zm/docs/ARCHITECTURE.md) — technical reference,
subsystem walk-through, full CLI table
[docs/archive/](https://github.com/bitmosh/cerebra-classic/blob/claude/github-docs-cleanup-hc15zm/docs/archive/) — historical design specs, development
log, state reports
Citation
Cerebra (pre-dyson-sphere baseline). Tag v0.4.4-pre-dyson, 2026-06-21.
https://github.com/bitmosh/cerebra-classic/releases/tag/v0.4.4-pre-dyson