v2.17.0 — Sub-Agent Maintainability & GPT Prompt Variants

Consolidate duplicated sub-agent declaration boilerplate behind a single source
of truth and shared helpers, split the monolithic hashline_edit core into
cohesive modules, and extend GPT-optimized prompt variants to every builtin
sub-agent.

Added

• GPT prompt variants for all five builtins — Explore, Oracle, Librarian,
  General, and Reviewer each define a systemPromptByFamily.gpt variant
  selected via resolveSystemPromptBody() when the configured nested model
  classifies as GPT family (gpt-*, o1, o3, o4). The variants follow
  OpenAI's GPT-5.x prompting guidance: outcome-first and shorter than the
  defaults (~25–47% smaller), top-level # headings with <xml> tags
  delimiting semantic behavioural blocks (tool rules, search/review contracts,
  output specs, stop rules, citation policy), explicit anti-filler openers, and
  fewer redundant negative rules. Each variant preserves its default persona's
  contract — read-only/full-access guarantees, tool allowlists, the literal
  ## Findings / ### High / ## Verdict headings the host parses from
  Reviewer, no-fabrication and citation-precision rules, and verification-gate
  discipline. Non-GPT families continue to receive the default persona.
• src/sub-agents/builtin-metadata.ts — a single source of truth for each
  builtin's name, description, and routing. The five declarations spread
  the shared metadata objects, and SUB_AGENTS in
  src/config/resource-metadata.ts is derived from
  BUILTIN_SUB_AGENT_METADATA, so the registry and the declarations can no
  longer diverge. A drift-prevention test asserts
  declarationToMeta(decl) deep-equals the derived SUB_AGENTS entry for every
  builtin.
• formatUserPrompt(main, context?, label?) in
  src/sub-agents/prompt-builder.ts — the shared user-prompt composer used by
  Explore, Oracle, General (default label) and Reviewer ("Review context").
• standardPrependOverlay(agentName) in src/sub-agents/runtime-overlay.ts
  — the shared prependSystemPrompt factory for read-only builtins and the
  YAML loader.
• Cohesive hashline_edit modules — schema.ts, anchors.ts,
  validate.ts, and result.ts split out of src/tools/hashline-edit/,
  keeping the applyHashlineEdits orchestrator, the canonical-path mutation
  queue, and tool registration in index.ts. The public surface
  (applyHashlineEdits, runQueuedHashlineEdit, safeRealpath,
  registerHashlineEditTool) is unchanged.

Changed

• The Explore default persona describes search breadth in terms of the caller's
  stated need rather than a thoroughness control, matching the
  delegate_explore parameters schema (question + optional context).
• The Librarian GPT variant uses brace-style URL placeholders
  ({owner}/{repo}/{sha}) in its citation template so the only angle-bracket
  constructs in the prompt are its real <xml> semantic blocks.

Robustness

• A runtime-overlay test asserts buildSubAgentRuntimeOverlay always renders
  the ### Current Date and ### Working Environment headings in order, which
  the Librarian and General personas reference via "see the runtime overlay
  above".