Analytics Phase 1 — make the existing numbers honest

[SaurabhPandey9752]

Phase 1 of the analytics rebuild brief. Backend only — no UI changes. The goal is
that the numbers already on screen stop describing us, so Phase 4 isn't a
better-looking view of the same bad data.

What was wrong

1. Self-referral attribution. inferSourceMedium() treated only the current
   request host as internal, so every other Taskly domain scored as an external
   referrer. We move visitors between our own domains deliberately (geo routing
   .ai → .ca, the .ai gateway hand-off, the legacy taskly.ca redirect) and
   each hop carries a Taskly Referer.
2. No internal-traffic exclusion. isDevHost() only blocks localhost. Nothing
   kept the team out of our own dashboards.
3. GMV destroyed by coupons. finalize-checkout wrote total_price = what Stripe charged, but total_price is the agreed deal.

Changes

Area	Change
lib/attribution/owned-hosts.ts (new)	Single source of truth for every host we own, derived from REGIONS + the legacy domain, apex and www.
lib/attribution/index.ts	inferSourceMedium() returns {null, null} for any owned host — same as direct. Deliberately not an 'internal' source, which would move the pollution rather than remove it.
lib/analytics/audience.ts (new)	One vocabulary for "whose traffic is this", the audience equivalent of date-range.ts. Default excludes team + bots; including them is an explicit ?includeInternal=true.
Migration 187	Reclassifies historical self-referrals across customer_profiles, tasks, bookings. Every changed row is recorded in a new attribution_reclassified audit table first, and customer_profiles also keeps the original inside its attribution jsonb. Reverses with one UPDATE … FROM (included as a comment).
Migration 188	profiles.is_internal + partial index + seeds the team/test email domains.
Migration 189	funnel_events.is_bot; rebuilds funnel_step_stats with p_exclude_internal / p_exclude_bots. Both default, so the currently-deployed 4-arg call keeps resolving during the deploy window. Uses a LEFT join to profiles — an inner join would have silently dropped every anonymous event, which is ~70% of them.
/api/track	Flags crawlers via the existing isBotUserAgent instead of dropping them, so "how much of this is bots?" becomes answerable.
/api/admin/{funnels,journeys,utm}	Honour the audience filter and echo it back so the UI can state what it's showing. Journeys filters internal ids in JS on purpose: user_id not in (…) is NULL for anonymous rows and PostgREST would have dropped them all.
finalize-checkout + process-event	total_price stays the agreed deal; the coupon lives in discount_amount; service_fee is Taskly's cut of the deal, not a negative number.

lib/analytics/leaks.ts is deliberately unchanged — it's a pure ranker over
data the funnels API has already filtered, so it inherits the exclusion. Adding a
parameter there would have been a no-op.

Acceptance — measured against production, read-only

Attribution, once 187 runs:

Table	Attributed now	Self-referrals cleared	Attributed after
customer_profiles	15	11	4
tasks	15	14	1
bookings	4	3	1

Internal exclusion, last 30 days (34 seeded internal accounts):

	Now	After
Sessions	1,528	1,443
Events	7,191	6,420

Verification SQL

with owned(host) as (values
  ('tasklyanything.ca'),('www.tasklyanything.ca'),('tasklyanything.in'),('www.tasklyanything.in'),
  ('tasklyanything.ai'),('www.tasklyanything.ai'),('taskly.ca'),('www.taskly.ca'))
select count(*) filter (where utm_source is not null) as attributed_now,
       count(*) filter (where lower(utm_source) in (select host from owned)) as will_clear,
       count(*) filter (where utm_source is not null
                          and lower(utm_source) not in (select host from owned)) as attributed_after
from customer_profiles;

Deliberately NOT done

The 7 broken GMV rows are not rewritten. The brief assumed they could be
recovered from the Stripe payment intent — they can't: all 7 have
stripe_payment_intent_id = null because a 100%-off coupon means Stripe never
creates one. They're arithmetically recoverable (discount_amount equals the full
deal on 6 of 7, and provider_payout / 0.8 on all 7), but that's rewriting money
rows on inference. Needs a founder decision — see the question in the thread.
The write-path fix means no new rows join them.

Gates

typecheck clean · 1,496 tests pass (2 new suites, 20 cases) · lint 0 errors
(214 pre-existing warnings) · knip clean · check:migrations OK

Before merge

1. Apply 187 → 188 → 189 to production first, then merge. The code reads
   is_internal / is_bot, and a missing column throws into a catch that
   returns [] — a silent empty dashboard.
2. Phase 0 (GA4 console) is still outstanding and is human-only: configure owned
   domains, unwanted referrals, internal traffic definition + data filter, and a
   hostName = localhost exclusion.

🤖 Generated with Claude Code