refactor(verify): add read-only mode, widen trigger, extract citation anchors

[bensonwong]

Summary

• Widen skill trigger: description now fires on any PDF, office doc (DOCX/XLSX/PPTX/etc.), image, scan, or URL task — not just /verify. Adds explicit "Tool precedence" section listing forbidden alternatives (Read for PDFs, pdfplumber, python-docx, Tesseract, etc.) so agents default to deepcitation prepare for all document work.
• Read-only mode: §1 Orient now picks between two modes — read-only (user wants text extracted, no citations needed) and verify (claims to check). Read-only gets a fast path at the end of §2 that stops after prepare + content delivery, skipping the full citation pipeline. Prevents agents from bailing to Read/pdfplumber when the task is just "summarize this PDF."
• Extract citation anchors to rules file: ~150 lines of truncation strategies, anti-patterns, Domain A/B explanation, and the 9-entry failure-modes table moved to rules/citation-anchors.md. SKILL.md now holds concise rules + pointer; the heavy reference is lazy-loaded only when needed. Consistent with existing rules/auth.md and rules/cloud-sandbox-constraints.md pattern.
• New CLI flags: --claim and --model added to verify command examples, with guidance on how to choose a claim label.
• Improved results summary: final line uses emoji counts (✅ N verified ⚠️ N partial ❌ N not found) instead of the old prose form.
• README: simplified install section — removes platform-specific subsections (CLI, Cowork, other agents), keeps npx skills add primary path + zip upload fallback.

Test plan

• Read-only path: ask "what does this PDF say?" — confirm skill runs prepare, reads .deepcitation/*.txt, returns content without citation markers, no verify call
• Verify path: ask /verify [claim] [doc] — confirm full §1→§4 pipeline runs as before
• Tool precedence: ask to "parse this DOCX" — confirm skill fires (new description matches) and routes through prepare
• Citation anchors: confirm rules/citation-anchors.md is referenced from SKILL.md and contains the full truncation/anti-pattern content
• --claim/--model flags appear in the verify command output
• README renders cleanly with two install paths

[claude]

PR Review

Overall this is a solid, well-motivated refactor. Extraction to rules/citation-anchors.md and the read-only fast path are both clear wins. A few issues worth addressing before merge.

ISSUES

1. README structure is broken

The new headings make Agent Skills a sibling of Install with nothing under Install, leaving an empty section. The zip upload numbered list is orphaned by a bare "Or" fragment, and the deleted step 4 was the only line telling readers what to do after uploading. Either consolidate the headings or restore that payoff sentence.

2. With-other-agents section removed without replacement

The old README noted that SKILL.md works with any agent that supports Markdown instructions. Still true and still in AGENTS.md, but gone from the public README. A single line under the zip path would restore discoverability without bloat.

3. Trigger description mixes metadata with instruction prose

The description frontmatter is consumed by skill loaders for trigger matching. The sentence "Preferred over Read, pdfplumber, python-docx, openpyxl, and Tesseract..." is instruction content, not a trigger condition. The Tool Precedence section in the body covers this adequately — trim the description back to trigger conditions only to avoid confusing loaders that do strict semantic matching.

4. Read-only resume assumption is too broad

The fast-path follow-up note says the .deepcitation output file is still valid without re-running prepare. This is only safe within the same session on the same document. Add a qualifier: only if the same document was used and the output file still exists. Otherwise agents may silently skip prepare on stale or absent output.

5. --model flag underspecified

"Your own model name as a human-readable string" is ambiguous — model ID (e.g. claude-sonnet-4-6) or display label? A concrete example and a note to omit the flag when the agent cannot determine its identifier would close the gap.

MINOR NITS

• The citation-anchors.md trigger list in SKILL.md omits one condition from the file's own "When to read this file" preamble: an anchor that "feels wrong". Keep them in sync.
• Test plan is missing coverage for --model, the emoji summary format, and the read-only-then-verify session resume flow.
• citation-anchors.md references paths under packages/deepcitation/docs/agents/ that appear to live outside this repo. A note that these are in a separate package would help agents installing the skill standalone.

WHAT WORKS WELL

• Extracting ~150 lines to rules/citation-anchors.md is the right structural move. SKILL.md is meaningfully shorter and the heavy reference lazy-loads only when needed, consistent with rules/auth.md.
• The read-only fast path fills a real gap. The "Why this branch exists" callout preempts the exact failure mode it is fixing.
• The --claim guidance (question vs. filename vs. URL vs. pasted content) is practical and well-differentiated.
• Emoji counts in the results line are a clear improvement over the old prose summary.
• The triage table row for read-only mode integrates cleanly with the existing table structure.