analysis: Phase 1 cohort query the shipped helper cannot express

[JasonYeYuhe]

Why

anonymous_activation_summary(p_days) groups by day, channel and selects no app_version at all, and applies no maturity window — it counts an install that first reported ten minutes ago as a failed activation. Phase 1 cannot be read through it.

Checked against the live function body (pg_get_functiondef, 2026-08-12): production matches the migration byte for byte. This is a limitation of the helper, not drift.

Read-only. No migration number consumed, no function created, no schema change.

What it encodes

Cohorts by first_seen_at, never by app_version — the upsert does app_version = excluded.app_version, so a 1.45 install that later updates reports as 1.47. Version columns survive as labelled diagnostics only.

Four things that are cheap to record now and expensive to rediscover later:

Metric is menu-open → provider	The disclosure card lives in the lazily-built MenuBarExtra and gates both events via hasSeenDisclosure. An install that never opens the menu is invisible, not a counted failure. Never call it "activation rate" unqualified.
iOS contributes nothing	AnonymousTelemetryCoordinator is constructed only in CLIPulseBarApp; CLI Pulse Bar iOS has zero references. The iOS 1.47 approval date is irrelevant to this table.
Deleting a row does not reset the client	installReported / activationReported are UserDefaults latches set only on a 2xx. A deleted install is absent from every future read. pg_stat_user_tables currently reads 7 inserts / 5 deletes / 2 live rows.
Construct validity	providers.isEmpty == false can be satisfied by cloud-synced rows from another Mac. It means "had a number to show", not "detected a CLI here".

Guard verification

Per the repo rule that a new guard must be shown to fire, not merely to pass — §4e feeds synthetic rows through the same predicates and asserts each control returns ≥ 1.

That caught a real bug in §4b: the naive

string_to_array(v, '.')::int[] < array[1,47,0]

reports "1.47" as pre-1.47, because a shorter array sorts before its own prefix. Confirmed against a probe of 1.45.0 / 1.47.0 / 1.47 / 1.46.9 / 1.48.0 / 2.0 / 1.47.1. The committed version pads to three components.

Verification run (production, read-only, 2026-08-12)

Every section executed against production. Backend CI gates run locally, exit codes captured to file rather than piped:

• ci_check_search_path.py · ci_check_rpc_contract.py · ci_check_date_windows.py (80 SQL files scanned, includes this one) · ci_check_alert_types.py · ci_check_user_id_cascade.py — all exit=0
• scripts/check_migration_numbers.sh — exit=0, 73 migrations, all unique

Current output — §1 attributes the empty cohort rather than just reporting zero:

stage	rows
all rows in table	2
channel recognised	2
first_seen_at ≥ channel 1.47 availability	0
AND matured ≥ 7 days	0

§3 computes the escape-hatch deadline as 2026-09-09, 27 days remaining.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Adds a read-only Supabase analysis SQL script to compute and sanity-check the Phase 1 “menu-open → provider detected” cohort funnel from public.anonymous_installs, including maturity gating, channel cutover dates, and negative controls intended to catch silent undercounting.

Changes:

• Introduces a staged attrition readout that explains why the eligible cohort can be empty (channel recognition, cutover, maturity, activation).
• Adds the main cohort rollup query with an explicit minimum-N decision gate for interpreting percentages.
• Adds an “escape-hatch” deadline calculation plus negative controls (including synthetic proof that controls can fire).

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.