v1.0.5 — Post-V2 doc hot-fix

Post-V2 doc hot-fix wave: six doc-only fixes from the V2 MoPla Snapshot
validation retro. Skill-content + plugin-manifest only release; no
@formtrieb/cdf-core / @formtrieb/cdf-mcp version bumps.

Fixed

• A1 — both SKILL.md files gain a §0.5 · Read-Path Resolution
  anchor clarifying that Read / yq / jq paths in the skill and
  its references/ + shared/ docs resolve relative to the
  SKILL.md base, not relative to the user's cwd. Eliminates the
  ls-discovery friction observed in V2 first-touch runs and gives a
  one-shot recovery instruction (find ~/.claude/plugins -name SKILL.md -path "*<skill-name>*") when the harness doesn't auto-
  resolve the relative path.
• A2-doc — synthesis.md §0 (Inputs) appends a Large-file
  fallback (>25k Read tokens) paragraph cross-referencing the yq+jq
  recipe in tool-leverage.md §4. Walker outputs exceeding the
  25,000-token Read limit (typical for DSes with ≥150
  component_sets) no longer block synthesis: the contract holds, only
  the read-strategy changes. Probe walker file size first
  (ls -lh .cdf-cache/phase-1-output.yaml); branch directly to
  fallback if >2 MB.
• A3 — tool-leverage.md §4 adds a §4.2 YAML extraction
  (mikefarah yq + jq) sub-section documenting the
  yq -o=json '...' file.yaml | jq '...' pipe pattern. Mikefarah yq
  rejects inline jq-style object construction (yq '.foo | {bar: .baz}'); the pipe pattern is the canonical fallback. Includes
  WRONG / RIGHT examples.
• A4 — walker-invocation.md §2 documents the
  total ↔ indexed_count field-name drift between what the walker
  emits (ds_inventory.component_sets.total) and what the snapshot
  schema expects (indexed_count). Both refer to the same metric
  (union of tree-resolved + remote-only sets); the alias-mapping is
  authoritative until the walker rename ships in v1.8.0.
• A5-doc — synthesis.md §2.7 (theming) appends a Fallback when
  theming_matrix.collections: [] is empty recipe. When the walker
  emits an empty collections list (typical when .cdf.config.yaml
  has no explicit resolver: block but regime: tokens-studio is
  set and tokens/$themes.json exists on disk), operators can derive
  theming-modifiers via a single jq -r 'group_by(.group) | …'
  invocation against tokens/$themes.json. Cross-validation note
  against component-side VARIANTS included.
• A6 — walker-invocation.md §2 adds an Output-Shape Examples
  block for jq/yq query construction. Several walker fields emit
  flat-string arrays (NOT object arrays) — standalone_components.*,
  pages.separators[], libraries.linked[], doc_frames_detected[].
  Documents the WRONG .utility[].name (returns null) vs RIGHT
  .utility[] query shapes.

Deferred to v1.8.0 (Synthesis-as-Code wave)

• A2-walker — cdf_extract_figma_file summary_only mode emitting
  phase-1-summary.yaml (drops inline propertyDefinitions for

  25k cases)

• A5-walker — walker auto-resolves theming_matrix.collections
  from tokens/$themes.json when regime: tokens-studio and no
  explicit resolver: is configured
• A7 — Renderer hyphen-line-break lint (cosmetic)
• B5 — libraries.linked walker output bug — under fresh-chat
  investigation in v1.8.0

Accepted as platform constant

• A8 — ToolSearch round-trip latency for first-time MCP tool
  invocation. Documented in v1.0.4's tool-leverage.md §1
  deferred-tool note; no further action.

No package version bumps in v1.0.5

• @formtrieb/cdf-core@1.0.x and @formtrieb/cdf-mcp@1.7.x unchanged
• .mcp.json still pins @formtrieb/cdf-mcp@^1.7.2
• Skill-content + plugin-manifest only release

Upgrade for existing installs

claude plugin marketplace update cdf
claude plugin update cdf@cdf
# Fully quit Claude Code (Cmd+Q) and relaunch — closing the window
# is not enough; the MCP launcher needs to reload .mcp.json env.