Make local agent memory usable with bounded V1 recall

[gnulnx]

Fresh sessions could miss projects, bury focused facts in a timeline, or receive a response truncated by the tool host. This implements the V1 product brief derived from #40: global project overviews, relevant factual recall, durable cross-session saves, and explicit tagged handoffs with bounded evidence.

• Add work_overview and exact, case-sensitive get_tag, plus blc overview and blc tag. Prefer notes/outcomes and spread overview evidence across projects and sessions.
• Bound complete retrieval JSON to 24,000 bytes, including provenance and coverage. Return one MCP text representation, meaningful excerpts, and explicit continuation/omission metadata. Keep large operator diagnostics behind blc index-status --details.
• Merge lexical and semantic candidates before pagination, retain immediate lexical recall while indexing, and prefer newer equally relevant authored statements. Agent provenance shares the same store.
• Retain automatic conversation content for seven days across producers. Keep authored notes/handoffs durable, remove expired vectors, and prevent reimport from renewing old content. Original provider transcript files remain intact.
• Update the installed skill, lifecycle guidance, and post-install examples around ordinary user workflows. Empty retained history can install successfully and demonstrate memory through new saves.

Validation: the full Linux suite with real CPU embeddings, systemd lifecycle, capture/restart, MCP stdio, and isolated product installation passed. Acceptance A–F covers uneven project traffic, three different changing facts, focused LiDAR evidence, exact tags, huge diagnostics/Unicode responses, and mixed agent provenance. A separate Python 3.11 wheel environment exercises the packaged product. Lint, distribution build, and strict package metadata checks pass.

Local review:

git fetch origin
git switch feat/usable-memory-v1
uv tool install --python 3.11 --editable . --force
blc install codex
blc doctor
blc overview --days 3

Start fresh Codex sessions to discover the new tools and skill, and review/trust Context handlers in /hooks if prompted. Ask for recent work; save a fact in one session and recall it from another repository; change it and verify the newest value; then tag current work with test-handoff and refresh that tag in another session. Exact prompts and automated test commands are in docs/DEVELOPMENT.md under V1 acceptance and local review.

Initial installation still backfills five recent sessions; trusted hooks capture subsequent work. Coverage explicitly reports limitations. This PR does not add Claude/Gemini integrations, cross-machine sync, configurable retention, or a server-side LLM. The MCP result shape intentionally changes to one JSON text block; the bundled client and health checks are updated together.

Closes #40.