fix: #5 — feat(claude): Claude Code adapter (Mode A + Mode B + skill)

[MiaoDX]

Closes #5.

Ships the two static assets the Claude Code adapter needs on top of the
settings.json snippet that scripture-mcp init --target=claude-code
already installs (landed in #4 / PR #14).

What's in this PR

• adapters/claude-code/output-styles/scripture-recap.md — style
  only, matches plan.md §5.1 verbatim. Sets
  keep-coding-instructions: true and instructs the model to treat
  <scripture_card> as a one-turn reflective frame. Contains no
  scripture text and names no specific tradition, per the
  issue's requirement.
• adapters/claude-code/skills/verse-inject/SKILL.md — the manual
  fallback skill for the slash-marker flow. Sets
  disable-model-invocation: true so the model never auto-triggers
  it. Body is 56 lines (< 60 per the issue), contains no scripture
  text, and instructs the agent to call mcp__scripture__lookup for
  every verse fetch — the preview shows display_ref, attribution,
  and checksum_sha256, then the user explicitly confirms before any
  injection happens.
• adapters/adapters.go — a tiny package that embeds the two
  static files via embed.FS. The package exists primarily so the
  test suite can assert structural properties of the assets without
  duplicating their content; the Claude*Path constants also
  document the canonical in-FS paths for the eventual installer in
  feat(install): one-line install.sh #7.
• internal/injector/envelope.go — the §6.3 envelope now carries
  a checksum: sha256:<hex> line so the Mode A preview surfaces all
  three integrity fields the issue requires: verse text + source +
  checksum. The model is instructed (in the envelope's leading
  paragraph) not to act on the checksum beyond reproducing it
  verbatim if asked.
• README.md — documents the symlink/copy install path for the
  two adapter assets. They're deliberately not auto-materialized
  by init: init's only acceptance criterion for feat(claude): Claude Code adapter (Mode A + Mode B + skill) #5 is recap on/off
  toggling, which feat(core): stdio MCP server + CLI subcommands #4 already met, and extending init to write files
  outside its existing splice surface would expand scope.

Acceptance-criteria mapping

Mode A — UserPromptExpansion hook:

• /bible John 3:16 shows a preview card with verse text + source + checksum (envelope now carries the checksum line)
• User must explicitly confirm before the model sees the verse (UserPromptExpansion shows the expanded prompt to the user before submission; the user hits Enter to confirm)
• On confirm, the next coding request is wrapped with the temporary envelope from plan.md §6.3
• keep-coding-instructions: true set on the active output style
• Works for all 4 marker traditions: bible, dao, sutra, quran (covered by feat(core): stdio MCP server + CLI subcommands #4's lookup-from-prompt regex)
• Hook exits silently if the prompt has no marker (covered by feat(core): stdio MCP server + CLI subcommands #4)

Mode A — verse-inject skill:

• adapters/claude-code/skills/verse-inject/SKILL.md exists and is 56 lines (< 60)
• disable-model-invocation: true in frontmatter
• Skill body contains no scripture text — calls mcp__scripture__lookup via MCP (asserted by TestVerseInjectSkillIsManualOnlyAndShortAndScriptureFree, which substring-scans the body against every bundled verse > 12 bytes)
• Manually invoking the skill triggers the same preview/confirm/inject flow (skill body steps 1–5)

Mode B — output style + Stop hook:

• adapters/claude-code/output-styles/scripture-recap.md defines style only — no scripture text or specific traditions (asserted by TestClaudeOutputStyleHasRequiredFrontmatter)
• Stop hook calls scripture-mcp recap --terminal (covered by feat(core): stdio MCP server + CLI subcommands #4 via init --target=claude-code --recap=on)
• Mode B toggleable via scripture-mcp init --target=claude-code --recap=on|off (covered by feat(core): stdio MCP server + CLI subcommands #4)
• Recap text does NOT appear in the next prompt's input — architectural property: Stop-hook stdout flows to the terminal, not into model_call. Rigorous end-to-end verification of this invariant is the explicit gate of test(critical): injection lifecycle + coding-quality regression #8 (test(critical): injection lifecycle); this issue treats it as a smoke-tested architectural guarantee.

Tests

• adapters/adapters_test.go — embedded-asset frontmatter assertions,
  line-count check on SKILL.md, no-scripture-text scan against every
  bundled verse, no-tradition-leak check on the style file.
• internal/injector/envelope_test.go — added the
  checksum: sha256:<hex> line to the envelope assertion list.

Verification

• make build clean
• staticcheck ./... clean
• go test ./... — green across all 7 packages
• python3 scripts/verify_quotes.py — 31,183 verses, 0 mismatches
• Smoke: echo '/bible John 3:16 Refactor X.' | ./bin/scripture-mcp lookup-from-prompt emits an additionalContext envelope containing the verse, the John 3:16 display ref, the KJV attribution, and checksum: sha256:8473c0b1c7664945528317faf77351258eb79f8b11ba821ef76d7e916cde711a.

---

Generated by Claude Code