v0.3.1 — assumption-layer conflict detection

What ships

Phase 0 reference implementation grew from "memory hub only" (v0.1) into a
four-layer substrate:

1. Memory hub — FastAPI + SQLite (WAL + FTS5) + JSON-RPC MCP bridge.
2. Mirror consolidation — nightly script that proposes merges /
   supersedes / link / surface-conflict ops; never auto-applied.
3. Assumption-layer conflict checker (v0.3) — Brian Williams /
   ATMS-inspired. Records optionally carry an `assumptions` list;
   conflicts are detected at the assumption-entailment level rather than
   embedding similarity. Catches the classic 'use account A' vs 'use
   account B' shape that textual dedup misses.
4. Weekly digest — plain-text Sunday email summarising what the
   system learned, so improvement is visible rather than invisible.

Plus a one-command AWS deploy script, a drop-in CLAUDE.md template that
makes agents auto-use the hub without prompting, and an INSTALL.md for
production-grade Cloudflare Tunnel + Access.

19/19 tests pass. Apache-2.0. Backward-compatible migration from v0.2
databases.

Companion repo

Architecture / safety / quickstart / philosophy:
https://github.com/starshard-ai/architecture-v1

Spec rationale

Every spec ambiguity encountered while building from the architecture
docs is logged in `SPEC-GAPS.md` (12 entries). Feeds back into the
architecture-v2 evolution.