0.5.10

New

• ails check <capability> lists targets with per-item scores; ails check <capability> <name> produces a focused report on one skill / rule / agent / etc. Capability
  vocabulary is read from the detected agent's config.
• Capability-level read-out re-introduced as Level: L# <Label> in the scorecard. Re-aligned to the canonical ladder (System / Primer / Composite / Scoped / Delegated /
  Abstracted / Governed / Adaptive, L0–L7). Detection is cumulative — display is the highest level where all prior levels also pass. Read-out only; rule applicability
  unchanged.
• Per-item health bars in capability listings — one bar per item, sorted worst-first, with (N: Xe/Yw/Zi) finding-count breakdown per row. Severity bands
  (red/yellow/green) chunked with blank lines.
• Top rules (by finding count) block in the whole-repo scorecard; top_rules + capability focus envelope in -f json.
• New rule CORE:S:0056 broken-markdown-link — validates [text](path) and [ref]: path targets resolve in freeform markdown.
• Opt-in generic file class via Markdown link-reachability (generic_scanning: true) — classifier BFS-walks outgoing links and tags reached in-tree .md files as
  generic, cycle-safe, depth-bounded.
• Per-agent memory entry locator covering Claude (~/.claude/projects/*/memory/, project + local agent-memory dirs) and Gemini (~/.gemini/tmp/*/memory/). ails check memory / ails check memories work end-to-end.

Changed

• One unified check display. Capability args narrow the input subset; the standard renderer prints the same shape with fewer rows.
• import-targets-resolve, import-depth-within-limit, broken-markdown-link raised to severity high — broken includes/links are functional context gaps, not
  stylistic warnings.
• Score bars share a consistent dim-gray empty baseline; only the colored fill varies.
• Surface-health row suppressed when only one surface has data (the top Score: already represents it).
• Typed PlatformUnavailableError replaces silent fall-through on non-JSON auth-platform responses.
• Explicit User-Agent: reporails-cli/<version> (auth) header on auth + platform requests for edge allow-listing.

Fixed

• Mechanical-check glob targets honor .ails/config.yml: exclude_dirs — rules with hard-coded path: globs no longer scan excluded directories.
• Mechanical violation attribution no longer falsely points to user-scope ~/.claude/CLAUDE.md when the project has no project-scope main. Skip rules +
  _classified_display gate the ~/ rendering on the classifier's precedence: user property (read from agent config patterns), not path-prefix heuristics.
• CORE:S:0056 + the link walker strip fenced code blocks and inline code spans before extracting links — [text](path) inside backticks no longer false-positives.
• Capability gates and L-level scoring no longer count user-scope ~/.claude/CLAUDE.md toward project file counts.
• Directory-glob patterns (trailing slash) in agent configs now enumerate *.md inside matched dirs; project + local agent-memory files classify correctly to
  subagent_memory.
• Mapper daemon stays attached across runs — socket-existence gate on PID, FIFO-only FD close (no more numpy/onnxruntime ImportError post-fork), SIGPIPE ignored. Warm
  ails check ~5.6s daemon-attached.
• walk_glob follows symlinked directories during descendant traversal with cycle protection.
• CORE:S:0024 + CORE:S:0056 now declare format: [freeform, frontmatter] so they fire on SKILL.md, agents/.md, rules/.md.
• Gemini memory locator updated to current upstream model (~/.gemini/tmp/*/memory/); legacy ## Gemini Added Memories section removed.
• ails check memory / memories enumerates ~/.claude/projects/<hash>/memory/ entries (glob path silently dropped ~/ patterns).

Removed

• framework/registry/levels.yml and its schema — dead code, never read at runtime. LEVEL_CAPS is hardcoded in source.

Full changelog: CHANGELOG.md