Implements all three open enhancement issues, evidence-driven and reversible, with text/--json/--md parity throughout.

Closes #7 — payload-quality axis (reap gap)

A second utilization axis: when an MCP tool fires, is the result signal or noise? usage.Stats now correlates each mcp__* tool_use with its tool_result and scores the payload (base64 blobs, data URIs, HTML soup, repeated boilerplate) into useful-vs-noise chars. reap gap adds a per-tool quality column and flags ⚑ noisy tools that fire often and return mostly noise — the class load-utilization marks green and mute never catches.

Live: get_viewport_screenshot (81×, ~316 KB base64) → 0% quality, flagged; execute_blender_code (478×, 99%) → correctly clean.

Closes #9 — usage-informed routing (reap route)

Opt-in, secondary to pruning. Organizes the skills that survive a prune into category routers driven by real firing evidence (not BM25/text similarity): frequently-fired skills stay exposed, the rare long tail is pushed behind leaf routers (by namespace, else dominant firing project) loaded on demand. Output is a plan — proposed, never auto-applied; never routes blind (no-evidence skills stay exposed). Leads with the "below ~150 skills native loading is enough" caveat; points at reap mute for the description-stripping half. Flags: --route-threshold, --route-min-skills.

Closes #8 — APM manifest (reap apm)

Read-only: turns a repo's firing evidence into a proposed Microsoft APM apm.yml (skills only, first cut). KEEP→include, REAP→omit, REVIEW→never auto-omit. --diff apm.yml reconciles: suggests adding fired-but-undeclared skills and dropping declared-but-cold ones (REVIEW protected). Upstream coordinates are recovered from apm.lock.yaml when present; otherwise a clearly-marked TODO comment, never an invented coordinate. Hand-emitted YAML (no new dependency).

Quality

Designed and built per-issue, then put through an adversarial multi-agent review (16 agents). Confirmed findings were fixed and locked with regression tests:

• apm: LoadAPMManifest no longer harvests phantom coordinates from comments/headers/URLs (was self-triggering bogus drops); encodeProject now matches Claude Code's exact [^A-Za-z0-9]→- bucket encoding (repos with ./_/space were silently emitting empty manifests).
• route: --route-min-skills skip now renders in every format (was leaking prose into --json/--md); deterministic category ordering.
• payload: flagged-noisy tools sort first so a single 0% anecdote can't outrank a high-volume offender.

Test plan

• go build ./..., go test ./..., go vet ./... all green
• New table tests for the noise heuristic, payload scoring, route tiers/categories/token accounting, APM selection/diff/coordinate harvest, plus the three regression tests above
• Live smoke of reap gap / reap route / reap apm (incl. --json/--md) against real ~/.claude transcripts

🤖 Generated with Claude Code