Improve MCP activation: proactive routing + verification failure analytics

[Q00]

Problem

PostHog (last 30d): 2,086 MCP-service users but only 807 reach any command_run — a 38.7% service → command conversion. Two gaps:

1. Hosts wait for the literal ooo syntax instead of routing suitable natural-language requests to Ouroboros entry tools.
2. When run/auto verification fails, we cannot see why (e.g. worker cd-ed into a subdirectory while the verify gate ran from root; concurrent workspace modification) — causes discovered only by manually reading user transcripts.

Additionally, #2278 removed mcp_serve_started / subagent_dispatch / install_started, breaking the attached → used funnel and the User-lifecycle dashboard series.

Plan

• Server instructions: outcome-first WHEN TO USE OUROBOROS routing (interview / start_auto / start_execute_seed / lateral_think), < 2KB, provider-neutral.
• Entry-tool descriptions: Use when / Result / Do not use when contracts; proactive lateral_think on stagnation signals.
• Closed verify_cause vocabulary on the deterministic verify gate, recorded on execution.verify.failed (+ local-only verify_cwd) and a privacy-bounded ac_verify_failed PostHog event.
• Reinstate the removed adoption events as daily-deduplicated rows (one per user/day/dimension), preserving fix(telemetry): reduce PostHog collection to core metrics #2278's volume reduction.

Acceptance criteria

See HANDOFF.md in PR #2295.

[ouroboros-agent]

There's already a PR addressing this issue: #2295. Review is in progress.

ouroboros-agent[bot]

[shaun0927]

Post-merge disposition — 2026-08-29

#2295 (ed2bc908) completed the source implementation for proactive MCP routing, verify-cause analytics, and lean adoption events. Keeping this issue open for the evidence boundary only:

• release a version containing feat: proactive MCP activation routing + verification failure analytics #2295;
• verify dashboard 1985514 insights for Service -> product command and First product command using interview, auto, run, pm, qa, seed, evaluate;
• record a full seven-day pre/post comparison by app_version;
• disposition directory-shaped nested artifacts in artifacts_missing_found_elsewhere as supported or explicitly excluded;
• move or reconcile the now-stale root HANDOFF.md implementation plan.

Do not claim activation lift before the cohort exists. No further routing implementation should start from this issue unless the measurement exposes a bounded defect.