v0.9.3 — Bug-fix release: restores two features that were silently broken in…

Bug-fix release: restores two features that were silently broken in Claude Code, plus CI/test/doc hardening.

Fixed

• Auto-context injection hook restored. The UserPromptSubmit hook imported context from longhand.cli, which exports only app after the cli.py → cli/ package split. The swallowed ImportError made the hook emit {} on every prompt — so automatic context injection had been silently dead since v0.9.0. (#11)
• MCP tools now load in Claude Code. 12 of 19 tools declared an outputSchema whose type was array/oneOf; Claude Code's MCP validator rejects any outputSchema.type that isn't "object", so those tools silently failed to load. Handlers return text (never structured content), so the schemas were decorative — they are now stripped before tools reach the client. (#9)

Changed

• SECURITY.md corrected to describe the two fixed-argv subprocess call sites and the single hardcoded-identifier PRAGMA f-string. The code was always injection-safe; the "no subprocess / zero f-string SQL" claims were stale. (#12)

Internal

• CI now tests Python 3.14 (non-blocking — onnxruntime's cp314 wheel intermittently raises an illegal-instruction SIGILL on GitHub's heterogeneous runners), gates coverage at --cov-fail-under=60, and runs a non-blocking mypy job. (#13)
• Behavioral test backfill for the CLI and the recall time parser; total coverage 66% → 73%, time_parser.py → 100%. (#14)