feat(bundle-analysis): JSONL-only gap analysis with full six-signal attribution

[colombod]

Summary

Implements a complete bundle usage gap analysis system. Answers "which named components of which bundles did I actually use versus what was declared?" across sessions and workspaces — using local JSONL only, no CI graph required.

What changed

Core library (context_intelligence/bundle_analysis/)

• Inventory scanner — three-tier schema (always_active / agent_level / mode_gated). Reads behaviors/*.yaml for always-active agents/context/recipes, modes/*.md contributes blocks for mode-gated components, agents/*.md frontmatter for agent-level tool declarations. Disk cache keyed on bundle {slug}-{hash} for near-instant re-runs. Supports both bundle.md and bundle.yaml formats.
• Signals reader — named sets throughout (not integer counts). Inventory used as reverse lookup table, enabling attribution for all six signal types including tools (via tool:pre events) and modes (via mode:activated / mode:changed events). Previously tools and modes were always-zero.
• Gap arithmetic — four gap types with specific component names: tree-shake, mode-refactor, config-gap, mode-never-activated. Two-tier evaluation: always_active and agent_level components get tree-shake/mode-refactor; mode_gated components get mode-never-activated.
• Orchestration order — inventory built first (scan_cache), then signals attributed using inventory as reverse lookup, then gap computed. Dropped AsyncCIClient requirement entirely.

Deleted

• GraphFetcher class and all 18 Cypher query files — no CI graph path remains.

Tooling

• tool-bundle-usage — removed CI client dependency, works JSONL-only.
• agents/bundle-usage-analyst.md — new agent contributed by bundle-usage mode. Calls bundle_usage once, interprets gap findings, writes structured markdown report.
• modes/bundle-usage.md — advertised: false mode that contributes the analyst agent and tool.

Tests

• 134 tests pass (was 116 before this work — 18 new tests for tool/mode attribution).
• ruff: all checks passed.
• pyright: 0 errors, 0 warnings.

Live verification

Against the bundle-use-inspectors workspace (real JSONL data):

context-intelligence:
  ✅ agents:  [graph-analyst, session-navigator, context-intelligence-design-facilitator]
  ✅ context: [context-intelligence-awareness.md]
  ✅ tools:   [bash, delegate, graph_query]

superpowers:
  ✅ modes:   [brainstorm, execute-plan, verify, write-plan]

Gap findings include tree-shake for 7 unused skills, mode-never-activated for the context-intelligence mode (never activated in this workspace), and 24 total mode-never-activated findings across all bundles — all based on real event data.

[bkrabach]

Reviewed for parent_id and Session B merge interaction: zero impact. This PR reads events.jsonl from disk only — doesn't touch the hook module, dispatch path, session-graph write path, or merge_state semantics. Additive surface area is justified by scope; 134 unit tests + live verification output in the PR body is solid coverage.

One mechanical note: this PR and #19 both touch behaviors/context-intelligence.yaml. Whoever merges second needs a manual rebase on that file. Recommend merge order #22 → #23 → #19, which puts the breaking change (#19) last.

Approved. The deletion of GraphFetcher + 18 Cypher query files is a one-way door — fine and consistent with the stated JSONL-first direction, just flagging.