feat(bank-sync): close the F2 report: gap backfill, consent and paused chip states, agent-triggered sync (#2165)

* fix(bank-sync): cron backfills the gap since the last successful sync

The daily incremental sync always asked the bank for the last 7 days. Any
pause longer than that (a lapsed subscription paid again, a consent renewed
after expiry, an outage) silently lost the days in between: the connection
came back, looked healthy, and the missing transactions never arrived.

The lookback now widens to cover the gap since last_synced_at plus one day
of overlap, capped at the 90-day PSD2 limit, and a gap of a month or more
asks for strategy=longest like the manual sync route does. Dedup via
external_id makes the overlap harmless. First syncs keep their 90-day path.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QdcDV7CngLkWUvfKSsxFhS

* feat(bank-sync): chip warns seven days before a bank consent expires

The transactions-page chip only reacted once a connection was already dead
(expired/error) or had gone stale. A consent that is about to end looked
healthy until the morning it stopped syncing. New "expiring" state when a
live connection's consent_expires is within seven days, the same threshold
as the consent-expiry email in the sync cron. Precedence: attention,
expiring, stale, healthy.

getChipState moves to lib/transactions/bank-sync-chip-state.ts so the
precedence is unit-tested; the component keeps the rendering only.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QdcDV7CngLkWUvfKSsxFhS

* feat(bank-sync): chip says paused when the subscription lapsed

The daily cron filters connections by the bank_sync capability, so a
company whose trial or subscription ended keeps status=active rows with a
frozen last_synced_at. The chip read that as "stale, check the connection",
which sends the user to re-authorise a connection that is perfectly alive.
56 of 191 active connections on prod were in this state on 2026-09-01.

New "paused" state, ranked above everything else, when the company lacks
bank_sync: hosted points at billing, self-host at the connector key, the
same split BankSyncNowButton already makes. getChipState takes an options
object so the clock stays out of render (react-hooks/purity).

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QdcDV7CngLkWUvfKSsxFhS

* feat(api): agent-triggerable bank sync in v1 and MCP

Closes the first wish in the F2 report: an integration could read bank
data but never refresh it. New POST /api/v1/companies/{id}/bank-connections/
{connectionId}/sync and MCP gnubok_sync_bank, both on a shared runner
(extensions/general/enable-banking/lib/trigger-sync.ts).

Cost is bounded structurally, not by policy: the window is never
caller-controlled (the cron's gap-aware 7 to 90 day lookback), a connection
synced within 15 minutes answers BANK_SYNC_COOLDOWN with next_allowed_at
(429 + Retry-After on v1; synced=false in-band on MCP so the agent reads on
instead of retrying), and a failing connection is throttled per process by
attempt time. A dead session is flipped to expired with a remediation that
hands the user the connect link: no API call revives a consent.

Gated on bank_sync like gnubok_connect_bank; scope transactions:write.
Registry, scope map, load-routes, spec snapshot and the generated
accounted-api skill updated; five BANK_SYNC_* / BANK_SESSION_EXPIRED codes
added to the structured-error registry. The web Synka-nu route is left as
is (see DECISIONS.md).

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QdcDV7CngLkWUvfKSsxFhS

* test(bank-sync): use the options object in the remaining chip-state calls

Four multi-line calls still passed the clock positionally after
getChipState moved to an options object; tsc flagged them (vitest did not,
the extra argument was ignored at runtime).

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QdcDV7CngLkWUvfKSsxFhS

* fix(api): address skeptic findings on the agent-triggered bank sync

Three refutations from the pre-publish skeptic pass:

1. Core imported the extension. The v1 sync route pulled the runner
   straight from @/extensions, which the core-build gate rejects and which
   left a live bank endpoint on zero-extension builds. The route now
   resolves it through the registry's services channel against a contract
   in lib/bank-sync/trigger-sync-contract.ts (same pattern as the
   Skatteverket read service) and answers EXTENSION_DISABLED when the
   extension is absent.

2. The idempotency cache stored the handler-level 429. A same-key retry
   after Retry-After, which is the documented retry, replayed the stale
   cooldown as a 400 for the cache's 24-hour TTL. withApiV1 no longer
   caches 429 responses; regression test added. The endpoint's pitfall no
   longer claims Idempotency-Key is mandatory (it was never enforced).

3. Two cron tests read the clock twice and failed whenever a millisecond
   passed between the reads. They now pin the clock with fake timers.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QdcDV7CngLkWUvfKSsxFhS

* fix(bank-sync): durable cooldown lease and review wording

Resolves the PR #2165 review findings in one pass.

Superagent P1: the attempt throttle was a process-local Map, so two agent
calls on different serverless instances (or a retry after a cold start on
a failing connection) could each bill an Enable Banking call, contradicting
the one-sync-per-15-minutes promise. New bank_connections.sync_lease_until
(migration 20260902150000), claimed with one conditional UPDATE before the
bank is called; Postgres row locking makes exactly one claimer win, the
rest answer BANK_SYNC_COOLDOWN. The lease stays for the full window on
success and failure. Tests cover the claim order, a failed attempt seen
from a second instance, a lost race, and an expired lease.

CodeRabbit: the =1 plural branch now reads "in 1 day" / "om 1 dag"
(daysUntilConsentExpiry rounds a partial day up, so "tomorrow" could be
today); the cooldown pitfall on the v1 endpoint, the MCP description and
the in-band cooldown instruction now say a cooldown can follow a failed
attempt and tell the agent to compare last_synced_at before deciding.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0125TMQQBjBBZG9YxP7wQWub

* fix(bank-sync): lease claim as a literal filter for the schema guard

CI's no-phantom-columns guard counts runtime-built query expressions and
its ceiling is exact; the templated `.or('sync_lease_until.is.null,...')`
claim added one. The column now defaults to epoch (NOT NULL), so "never
claimed" is just "expired long ago" and the atomic claim is a single
literal `.lte('sync_lease_until', now)` the guard can check. Migration is
unshipped (same PR), so it is edited in place.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0125TMQQBjBBZG9YxP7wQWub

* fix(bank-sync): runner verifies company membership before the lease

Superagent (round 3): the MCP path reached the shared runner without a
membership check of its own. Both callers do enforce it upstream
(withApiV1's company resolution and resolveMcpCompanyContext in the MCP
dispatcher), but the runner writes transactions and bills a bank call, so
it now checks company_members itself, before the cooldown and the lease
claim, and answers NOT_FOUND for a non-member. The viewer check that was
buried inside the sync block moves up with it.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0125TMQQBjBBZG9YxP7wQWub

---------

Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>

DECISIONS.md

@@ -1495,4 +1495,6 @@ One line per decision: `[YYYY-MM-DD] <decision>: <why>`. Appended by agents and
 [2026-09-02] Removed the skattekonto drift email (skattekonto.drift_detected event, handler, /api/extensions/skatteverket/skattekonto/drift route, cron hook) instead of fixing it: it alerted on raw saldo-vs-1630 gaps that unbooked rows explain by construction (2026-09-02: Arcim 35 842 kr, 100% explained, while the Hem notice and reconciliation page said nothing was wrong), repeated every 24 h, and was the only surface of a May-2026 feature whose promised dashboard tile was never built. Since 2026-08-25 the reconciliation page and the Hem notice (detectSkvUnexplained, gated on unexplained_difference) are the surface. Considered gating the mail on unexplained_difference + once per episode (built, then dropped): after that gate it only fires on integrity findings the engine itself calls 'never a user task'. skattekonto_drift_tolerance stays (Hem notice reads it); stale skattekonto_drift_last_alert_at rows in extension_data are inert.
 [2026-09-02] parties phase 0, golden set stays out of git: the labelling sample is prod voucher text with person names (salary, expense claims) and the repo is public, so the draw SQL is versioned but the rows and labels live in gitignored dev_docs/parties/golden/.
 [2026-09-02] MCP eager-auth flag (`auth=required`) on the claude.ai connector links instead of reverting lazy auth: claude.ai's two-step Add-custom-connector dialog probes the URL without credentials and pre-fills Authentication "None" when the lazy handshake answers 200, which blocks the sign-in later; per Anthropic's docs a 401 is the only answer it reads as OAuth. The flag lives in the URL, so the links we control (Settings, onboarding checklist, both docs pages, website) get OAuth detected while the bare URL keeps lazy auth for Claude Code, the plugin, Cursor and ChatGPT, and existing connector records stay untouched. Rejected: keying eager auth off `client=claude-connector` (documented as telemetry-only) and sniffing the probe's user agent (fragile, undocumented).
+[2026-09-02] Agent-triggerable bank sync shipped (v1 POST /bank-connections/{id}/sync + MCP gnubok_sync_bank), lifting the 2026-09-01 deferral: Emil chose to close every open F2 item in one PR. The cost worry is bounded structurally instead of by policy: the window is never caller-controlled (gap-aware 7 to 90 days, same helper as the cron), a connection synced within 15 minutes answers BANK_SYNC_COOLDOWN with next_allowed_at (MCP returns it in-band as synced=false so agents read on instead of retrying), and failures are throttled per process by attempt time. The web Synka-nu route is left untouched rather than refactored onto the shared runner: it carries UI-only behaviour (caller-chosen days_back up to 365, SIE sweep stamping) and a regression there would hit every user for a code-sharing win.
+[2026-09-02] Bank-sync cooldown is a durable lease column (bank_connections.sync_lease_until, migration 20260902150000) claimed with one conditional UPDATE, not a process-local attempt map: the security scan on PR #2165 showed the map is bypassed by a second serverless instance or a cold start, so two agent calls could each bill Enable Banking. A column add was chosen over reusing extension_data because PostgREST cannot express an atomic conditional upsert there; the nightly cron deliberately ignores the lease.
 [2026-09-02] Grok links carry auth=required like the claude.ai link (#2159), decided from a live test: on the lazy URL Grok's connector dialog listed all 150+ tools and never opened the sign-in, so it reads the 200 probe as an authless server exactly as claude.ai does. The flag lives in one helper (mcpServerUrl / sideDoorServerUrl in lib/onboarding/checklist.ts) so the settings row, the onboarding side door and the deep link cannot drift; ChatGPT stays lazy because its developer mode honours the 401 on the first protected call.