Guillem Arias 83c64b9e94 fix(goals): pledge lifecycle + connected-account detection ...

Behavioural fixes touching Goal, GoalPledge, the reconciler and the
goals controller. No schema change.

B5 — connected-account detection covered only Plaid. SimpleFIN, Brex,
Enable Banking, IBKR, Kraken, SnapTrade and Lunchflow users got
"manual_save" pledges by default; their auto-synced Transactions then
failed to match (reconciler matches Transactions to "transfer" pledges
only). Pledges sat in the yellow banner until expiry. Switch the
detection to !Account#manual?, which mirrors the existing
`Account.manual` scope (no account_providers, no plaid_account_id, no
simplefin_account_id). Add `Account#manual?` so the per-instance and
per-query checks can't drift.

B7 — `extend!` widens `expires_at` but `matches?` was anchored on
`created_at ± 5d`, so an extension that pushed the expiry past day 5
didn't actually buy any match runway. Widen the upper bound to
`max(created_at + 5d, expires_at)`. The lower bound stays at
`created_at − 5d`.

B8 — `Goal#open_pledges` returned `status: open` regardless of expiry.
Between a pledge timing out (day 7) and the 15-min sweep job marking
it `expired`, the show page rendered a ghost yellow banner with
"0 days left" that the reconciler would no longer touch. Add
`expires_at >= NOW` to the scope so the visible state matches the
match-eligible state.

B9 — Double-click on Record pledge produced two identical open
pledges, which then stacked as two yellow banners. Add a create-time
validation rejecting duplicates against (goal_id, account_id, amount,
status=open, expires_at >= NOW).

B10 — The reconciler used `transaction.with_lock` but didn't lock the
pledge. Two concurrent reconcile attempts on different transactions
could both target the same pledge; one would lose to the partial
unique index on `transactions.extra->'goal'->>'pledge_id'` and the
RecordNotUnique was caught by the outer StandardError rescue, which
silently dropped the other transaction's match attempt entirely.
Lock the pledge first, re-check `status_open?` inside the lock, and
catch RecordNotUnique alongside RecordInvalid/NotOpenError in the
reconciler — so on a lost race we fall through to the next candidate
pledge instead of exiting the loop. Extract the Valuation-match path
to `GoalPledge#resolve_with_valuation!` so it goes through the same
locked status-recheck.

B12 — When a goal is destroyed, `dependent: :destroy` reaped pledges
but left `transactions.extra["goal"]["pledge_id"]` pointing at the
now-deleted UUIDs. The partial unique index on that JSON path then
indexed stale references. Add a `before_destroy` on GoalPledge that
clears the matching transaction's `extra` if it still points back to
the pledge.

B6 — `last_matched_pledge_at` used `goal_pledges.maximum(:updated_at)`
on matched rows. Any backfill or sync-resync that touches a matched
pledge bumped `updated_at`, so a single resync set every goal's "Last
saved N days ago" header back to "today". Switch to the entry's
`date` via a join through `matched_transaction_id`, which reflects the
date the money actually moved.

B22 — `scope :chronological` ordered DESC, the opposite of what the
name promises. Rename to `:reverse_chronological` and update the one
caller in `goals#show`. (Other models' `chronological` scopes are
unrelated and ordered correctly.)

Also: preload `account_providers` on `linked_accounts` in the index
and show controllers so `Account#manual?` walks the in-memory
collection instead of triggering N queries.

Tests: add fixture-backed coverage for extend-widens-match-window,
post-extend rejection beyond expiry, and the duplicate-pledge
validation. Existing assertions still hold against the new
`matches?` window math.

2026-05-14 19:12:28 +02:00

..

reconciler.rb

fix(goals): pledge lifecycle + connected-account detection

2026-05-14 19:12:28 +02:00