refactor(verify): enforce bold=k identity, multi-line l arrays, hard brevity rules

[bensonwong]

Summary

• Bold = k identity: Bold text in the body and k in CITATION_DATA must now be identical verbatim phrases — the reader clicks the bold term and sees those exact words highlighted in the evidence popover. Previously these could diverge (different paraphrase vs. source phrase).
• Hard ≤4-word limit: Clarified this is a hard limit on both bold text and k, not a suggestion. Added concrete truncation examples showing how to reduce long evidence phrases to their 2–3 most distinctive words.
• Multi-line l arrays: l must now include the anchor's line plus 1–2 adjacent lines. A single-line l risks the anchor filling the entire evidence paragraph, which renders with no visible highlight.
• Layer 2/3 UX explanation: Rewrote the progressive-disclosure section to explain what actually happens at each layer — including that a full-paragraph k produces no highlight, and that very long anchors lose readability in the keyhole strip.
• Sub-agent prompt updated: The inline citation format instructions passed to sub-agents now reflect all three rules above.

Test plan

• Run /verify on a document with long evidence phrases — confirm bold terms are ≤4 words and match k exactly
• Confirm l arrays in generated CITATION_DATA include 2–3 lines, not a single line
• Verify the evidence popover highlights the bold term (not a different phrase)

[claude]

PR Review: refactor(verify)

Overall a well-reasoned refactor. The three-layer disclosure explanation is genuinely helpful for grounding the rules. A few issues worth addressing before merging:

Bug: Invariant uses wrong command name

The updated Invariants section says to "only run deepcitation auth if prepare or verify output contains the exact phrase 'action needed'" — but deepcitation auth is not the command used anywhere else in this skill. Every other reference uses deepcitation login (auth.md Step 1, the login block in Step 1). Looks like an accidental substitution during the refactor.

Inconsistency: --summary to --text change not fully reflected downstream

The prepare command changed from --summary to --text, but Step 2 still refers to "the summary" and "the summary JSON", and the CLI-version note still references deepTextPromptPortion as a fallback indicator. If --text produces plain text rather than structured JSON, the instructions for reading attachmentId and deepTextPages are potentially wrong. Worth clarifying whether --text still produces the same structured output and updating Step 2 references accordingly.

Stale reference in auth.md

The auth-failure block still says: "Do NOT generate citation markers (anchor)". Should be updated to the new bold term [N] format for consistency with the rest of the skill.

Minor: Sub-agent prompt verbosity

The sub-agent citation instructions are now one long dense paragraph vs. the original two-sentence bullet. Functionally fine for an LLM-facing prompt — just noting it if readability is revisited later.

Positives

• The three-layer progressive disclosure explanation (Scan / Highlight / Explore) maps cleanly to the hard rules and makes the "why" easy to internalize.
• Concrete truncation examples are far more useful than abstract word-count rules.
• The multi-line l requirement with the rendering explanation (single-line array causes no visible highlight) is a great addition.
• Pinning deepcitation@latest across all commands is the right call for a prompt-based skill.

The wrong command name in Invariants and the --summary/--text consistency gap are the two must-fix items. The auth.md stale reference is low-risk but easy to clean up.