fix(mcp): supplier-invoice-from-inbox resolves FX through the shared resolver, with the cache and an override (#2173)

* fix(mcp): supplier-invoice-from-inbox resolves FX through the shared resolver, with the cache and an override

MCP feedback seq 299742: eight USD supplier invoices staged from the inbox
in one batch, three got a Riksbanken rate and five came back
exchange_rate: null / exchange_rate_source "lookup_failed", reproducibly,
for ordinary weekdays in April-August. None of the five could be approved
(the executor refuses with SI_FX_RATE_MISSING; it never books 0 SEK), and
the tool offered no way to supply the rate.

Cause: the tool called fetchExchangeRate without the supabase client, so
neither the shared exchange_rates read-through cache nor the
last-cached-observation fallback was reachable. Riksbanken's IP limiter
answers 429 after about five requests in a burst (a weekend date costs
two: exact-date 204, then the 7-day range), sends no Retry-After header,
and asks for ~54 s, which the 5 s retry cap cannot honour. The pass/fail
split was request ordering, nothing about the dates.

- Resolve through resolveSupplierInvoiceExchangeRate with the client, the
  same resolver the commit executor and the v1/web write paths use, so the
  staging preview and the commit agree and the cache is consulted and
  warmed.
- New input exchange_rate_override (SEK per 1 unit of invoice currency),
  trusted verbatim like the web form and v1; validated positive and finite,
  refused as implausible past the resolver's bound, rejected on a SEK
  invoice. Source is echoed as "supplied".
- When the lookup still fails, the preview carries exchange_rate_hint
  saying approval will refuse and naming the override that unblocks it.

tools/list: +1 property (~25 tokens), ledger line added in
payload-size.bench.test.ts; ceiling unchanged. The retry cap is left as is:
waiting a minute inside a tool call or the sync cron's fan-out is a design
call, and the cached fallback now covers the common case.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013yw62FMXGSzo6icFDiBwP3

* docs(decisions): FX override and retry cap on the inbox supplier-invoice tool

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013yw62FMXGSzo6icFDiBwP3

---------

Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>

DECISIONS.md

@@ -1513,6 +1513,7 @@ One line per decision: `[YYYY-MM-DD] <decision>: <why>`. Appended by agents and
 [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.
+[2026-09-02] create_supplier_invoice_from_inbox gained exchange_rate_override (+~25 tokens against a tools/list budget at zero headroom) and the Riksbanken retry cap stayed at 5 s: a staged foreign invoice with no rate is unapprovable (SI_FX_RATE_MISSING) and the override is the only agent-side unblock, while honouring the limiter's 54 s ask would park a tool call or the sync cron's fan-out for a minute per 429; passing the supabase client to the shared resolver (cache + last-cached fallback) removes the common case instead.
 [2026-09-02] reconcile_match staging folds N:1 groups back together by journal_entry_id (links from the dry run that share a verifikat and carry no allocated_amount), mirroring the existing bank 1:N fold, instead of threading the caller's original pairs through: the dry run is the source of truth for what was validated, a verifikat can only be settled once so shared-JE links can only have come from one N:1 pair, and use_proposals has no caller pairs to thread. The remaining gap (dry run validates only pair shape for non-split pairs) is the engine's to close, not the tool's.
 [2026-09-02] parties children/roles reference parties(id, company_id) with composite FKs, not parties(id): a party UUID from another tenant is rejected by construction instead of relying on each writer to check; ON DELETE SET NULL (party_id) on customers/suppliers because a plain SET NULL would null company_id too (Superagent P2 on #2162)
 [2026-09-02] Party suggestions attach only by explicit party_id, org number or an exact ledger key already in alias_keys; same-core text is reported as similar_to for a person to decide and identities are withheld when a key mixes org numbers: the selection eval measured 9% false merges on trade names shared by distinct legal entities (Fortnox AB / Fortnox Finans), so text never merges