BoB v0.98.0

BoB v0.98.0 — the memory-hardening release

Added

• Single-writer memory write fence (OS-level, fail-closed). Every memory writer must hold a family-scoped lock (one per vector-store service + collection family, spanning all vector dims). Memory enabled with no armed fence refuses to boot; a second writer degrades to read-only with honest health (HTTP 423 memory_write_locked) instead of corrupting the store; permission problems report as permission problems.
• Experimental zero-Docker vector store (opt-in). The zvec embedded store runs as a supervised child process owned by core (crash → clean error, next call restarts it). Opt in via config/memory_stores.toml + pip install zvec==0.5.1; Docker Qdrant remains the default.
• BobLKS: a local corpus instance — ingest() / retrieve() over your own documents through the same fence, fingerprint, and provider seams as agent memory.
• Asymmetric embedder seam — typed embed_query / embed_doc with per-slot instruction templates (part of the embed fingerprint), batching, and a first-write dimension probe (writes fail closed on mismatch, reads fail open).
• Retrieval eval harness with an honest ceiling (bobclaw-core/evals/retrieval/): author-blind paraphrase set, Wilson CIs, paired outputs. Measured recall@10: qwen3-embedding-4b 65%, 0.6b 40%. Earlier 95%/88% figures were lexical-leakage-inflated and are relabeled as an embedding smoke.
• Recall backfill past dangling vectors — bounded overfetch/paging to reach top_k valid facts instead of silently truncating.
• CI + release-surface baseline runner — .github/workflows/tests.yml (Windows + Linux from the pinned locks) and ./run_baseline_tests.ps1, which exits non-zero if ANY suite fails.
• UPGRADE.md — the supported tag-to-tag upgrade runbook.

Changed

• Default embedder: granite-embedding-311m → qwen3-embedding-4b (768 → 2560 dims, last-token pooling), decided on the paraphrase eval above. granite remains the documented CPU-light alternative (not yet measured on the new eval set — an open follow-up). Existing memory stores must re-index — see UPGRADE.md.
• Compose project naming is pinned and fixed container names are removed, so two installs coexist on one host (distinct COMPOSE_PROJECT_NAME + the new BOBCLAW_*_PORT datastore-port overrides). Existing installs: see UPGRADE.md before your next compose up.
• L1 fact dedup identity excludes volatile turn metadata (extractor v1→v2), so a repeated fact no longer multiplies.

Fixed

• Qdrant container healthcheck was always failing (image ships no wget) — now probes /healthz over bash /dev/tcp.
• MEMORY_SLOTS_FILE was documented but ignored by bootstrap — a second install could not point its embedder/extractor slots away from the shipped defaults. Now honored (with a regression test).
• start-local.ps1 probed fixed ports for its idempotence check and could mistake a coexisting install's gateway for its own — now probes the ports the install is configured for.
• Memory integration smokes isolate to throwaway collections with a fail-closed write guard.

Known limitations (honest)

• zvec approximate search trails exact search on recall (21/40 vs 26/40 through the full production path); a rescoring tier is planned for v0.99 — Qdrant stays the recommended default.
• The document parser applies soft limits only (hard cap filed for v0.99).
• pip-audit reports UNAVAILABLE (never silently "clean") when not installed.

Full notes in CHANGELOG.md. Release gate: clean-clone, shipped-install-path baseline — core 2846 passed / 0 failed / 6 skipped, gateway / pipeline / KMM / pip-check green — plus a live production-shaped stand-up (login, chat, fan-out, memory write + recall, restart survival) on a second isolated install.