Skillscript

Markdown support + strict-target detection + # Autonomous: true
header. Closes the cold-author footgun where .skill.md files with
markdown prose around the skill code triggered misleading
missing-dependency cascade errors (fbf10206). Adds the canonical
declarative marker for autonomous-execution skills.

Spec: Perry approval 08a08316 + amendment f352413d + final
greenlight efad035f.

Added

• Markdown extraction at parser layer — parse() scans the
  source for the first ```skillscript or ```skill fenced block
  and parses its contents. Cold-author LLMs writing markdown prose
  around their skill code get extraction automatically. Lives in the
  parser, not the skill store — clean layering per Scott's read
  ("storage shouldn't be the format-dispatch layer").

  Lenient by design: if no fenced block is found, the whole source
  parses as raw (existing behavior). Backward-compatible with every
  existing .skill.md file — no migration, no breakage. The fenced-
  block convention is the recommended shape for files with prose;
  pure-code files keep working as-is.

  # Welcome
  Use this skill for the morning sweep.
  
  ```skillscript
  # Skill: morning-sweep
  # Status: Approved
  run:
      ! morning
  default: run
  \`\`\`
  

• Strict-target-detection — target declaration lines now require
  <ident>: shape (matching [A-Za-z_][\w-]*). Prose lines like
  ## Use this: or Note (important): are silently treated as
  comments instead of misread as malformed target declarations. Pairs
  with markdown extraction: even without a fenced block, prose lines
  no longer cascade into misleading missing-dep errors.

• # Autonomous: true | false header — declarative authorship
  intent marker for unattended-execution skills (cron-fired, agent-
  fired, etc.). Today silences unconfirmed-mutation lint warnings
  for the whole skill (since the user-confirmation pattern doesn't
  apply to autonomous skills). Implemented as a category marker on
  ParsedSkill.autonomous so future rules + scheduling defaults +
  discovery surfaces can hook into the same field without breaking-
  change — per Perry's framing in efad035f.

Fixed

• unconfirmed-mutation lint conditional on # Autonomous —
  pre-existing rule from v0.2.11 (Bug 6) now properly distinguishes
  interactive skills from autonomous ones. Cold-author skills that
  legitimately invoke mutating tools without ?? confirmation
  (cron-fired log-monitoring → YouTrack issue creation, etc.) declare
  intent via the header instead of seeing false-positive warnings.

Implementation notes

• Parser-layer extraction matters because file-extension dispatch
  (the rejected alternative path) would have coupled the skill store
  to markdown semantics. Storage stays format-agnostic; parser handles
  the markdown wrapper concern locally. No skill store changes in this
  release.

• Tests: 23 new across tests/v0.4.2-autonomous.test.ts (10 —
  header recognition + lint conditional + help-content) and
  tests/v0.4.2-markdown.test.ts (13 — extractor + parse integration +
  strict-target-detection + end-to-end cold-author footgun closure).
  Total 965 passing in the suite (3 long-skip browser dogfood).

• No LOC ceiling nudge. 6593/6600 — under by 7. The ergonomic
  fixes are small enough to fit in the v0.4.1 ceiling without
  expansion.