v1.14.0

Fixes from the first real-environment evaluation: the Skill installed in Claude Code, screenshots organised end to end
by fresh Opus and Sonnet agents that read only the Skill, blind dedupe and answer suites, and the trigger suite in real
claude -p sessions. Data stays V1/V2; no migration is needed.

Evaluation results: dedupe retrieval recall@10 1.00; blind dedupe judgments 0 % false merges and 3.4 % missed merges
(Opus, Sonnet); blind answers 1.7 / 2 from a separate judge with no unsupported claims; triggers pass with Sonnet
(recall 1.00), while Haiku's recall (0.73–0.80) stays below the gate. See evals/README.md.

Added

• Dedupe review sheet: every dedupe task writes review-sheet.md, incoming questions grouped by report topic with their
  candidates and the bank's existing questions of the same topic, so paraphrases with no shared wording are compared.
• Short refs (n1… incoming, e1… existing) accepted in decisions; default_action: "KEEP_DISTINCT" covers unlisted
  questions; ingest submit --top-k; dedupe --decisions (alias of --input).
• --defer-review on ingest finalize and dedupe: REVIEW items no longer block the whole stage; they stay separate
  questions and wait for the user in the Web reader (合并裁决). dedupe --resolve applies decisions on a committed run's
  deferred items as a new stage; undecided items stay queued.
• page-text <url>…: saves the readable text of cited pages under cache/pages/ and returns a file for
  answer --page-texts, for hosts whose web tool returns summaries (the second command that goes online); pages that
  redirected elsewhere are flagged.
• Extraction rules for phone screenshots: overlapping scrolls, questions cut across images, very tall images, the same
  interview posted twice, abbreviations as company evidence and platform evidence.
• Dedupe judgment table (generic vs implementation, definition vs countermeasure, supersets, languages, project questions,
  follow-ups vs siblings) with examples that are not in the evaluation set.
• Role track ai-agent.application (AI 应用开发); technologies agent-skills, claude-code, codex, cursor,
  github-copilot, ragas (catalog 1.2.0).
• Rule for “the N most frequent” when place N is tied.
• tools/run_trigger_eval.py --max-turns; run directories record the Skill version.

Changed

• A merge or related target may be any active question of the bank or stage, not only a retrieved candidate; such
  targets are audited as outside_candidates, and unknown targets fail with the ID named. Committed questions can be
  linked or merged later with dedupe-candidates --question.
• dedupe-candidates returns the same compact shape as ingest submit (counts, review sheet path, a 15-item preview with
  refs and scores), not the task file; run-show pages the full task. Output trimming also reaches lists nested one
  level down, so a 400-question ingest submit no longer prints 456 KB.
• Dedupe tasks store only ID and wording for candidates (a 400-question import wrote 4.5 MB of task files).
• Candidate ranking keeps the real score for same-topic questions instead of flooring them to one value, so the top-k
  cut among them is no longer arbitrary.
• ingest finalize --task refuses a decisions file that names another task.
• Report topics are ordered by question count.
• SKILL.md description lists more Chinese trigger phrases and excludes self-introductions, salary negotiation and
  post-interview emails. Rarely needed detail moved out of the end-to-end reading set (export formats, answer recheck,
  merge audit), which stays under 30K characters.
• The Python version error explains how to find a newer interpreter.

Fixed

• tools/run_trigger_eval.py counted API errors (for example an unresolvable model alias) as “not triggered” and aborted
  on harmless CLI warnings; it now judges each case by the session's own result events.