Release 0.15.0: reviewed writes and CLI state integrity

[renezander030]

What & why

Reviewed task writes could execute twice under concurrent apply, use a payload changed after approval, or overwrite a task changed while awaiting review. State import previews also accepted invalid restore data, and CLI reads could appear successful despite incomplete results.

This prepares 0.15.0 with:

• State restore preflight: recognized JSON/JSONL shapes, schema versions and optional export checksums validated before the first write, including dry-run.
• Automatic target revision checks before applying reviewed task writes.
• Durable exclusive apply claims, with interrupted and uncertain outcomes retained for inspection.
• Boolean/value argument contracts and literal arguments after --.
• Approval digests bound to the staged kind and payload.
• Per-probe doctor deadlines and truthful degraded reports.
• Opt-in --require-complete read exits with JSON output preserved.
• Corpus cache isolation by adapter, working directory and source configuration.
• PID/host lock ownership and cleanup that preserves replacement locks.

All public npm versions, internal dependency ranges, the lockfile and MCP registry manifest are aligned at 0.15.0. The release guide covers publication order and verification.

Validation

• npm test: 472 tests passed, plus lint, disclosure/claims checks and all repository proof commands.
• npm run check:publish and npm run check:release passed.
• mcp-publisher validate server.json passed.
• Clean installation from all 13 public package tarballs: CLI version, strict find and doctor; adapter/core imports; packaged MCP stdio initialization and tools/list.

Review notes

• Based on main eb1da9d. The changes cover Core persistence and the CLI; adapter implementations and the MCP command surface are unchanged.
• Older approved task-write items without a payload digest must be staged and approved again. Interrupted applying or failed items require inspecting the backend before a fresh proposal.
• Revision checks remain optimistic on backends without native conditional writes. Restore preflight prevents invalid-input partial restores; filesystem failures during later writes remain possible.
• README edits may overlap textually with bench: reproducible 50-question recall comparison with Engram #39; benchmark code and data from that PR are untouched.
• Publication is held until maintainer acceptance of the final PR commit. Intended destinations: public npm workspaces, the existing MCP registry entry, and one matching GitHub Release.

Type

• Bug fix
• Feature / enhancement
• Docs

Checklist

• npm test passes
• Examples and fixtures use generic data
• Core changes checked through the full adapter and MCP suites
• No new adapter