fix(invoice-inbox): read whole PDFs (last-page slice + truncation retry) (#2014)

* fix(invoice-inbox): read whole PDFs (last-page slice + truncation retry)

PDF extraction read only part of well-structured PDFs, two confirmed
mechanisms (21-day prod window: 49 sliced docs, 29 silent empties):

- The auto-extract page budget was 3 (Bedrock-latency legacy, issue #553)
  and the slice kept only the first pages, so multi-page invoices lost the
  final page where totals, OCR and 'Att betala' sit. The budget is now 8 on
  pdf-native backends (Claude reads PDFs directly); the slice always keeps
  the last page. Rasterizing self-host backends keep the old budget of 3.
- A max_tokens-truncated model answer was parsed as-is, failed, and became
  an all-null extraction with no trace. extractFromDocument now reports
  stop_reason max_tokens / finish_reason length as truncated; the extractor
  retries once at double AI_EXTRACTION_MAX_TOKENS and logs
  ai_extraction_truncated either way.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012xosyW53HUa9JoFiDayhSk

* fix(invoice-inbox): sweep cutoff covers the slower two-call extraction

Skeptic finding on #2014: the crash-recovery sweep flipped 'processing'
rows to an empty skeleton after 2 minutes, but a deferred extraction can
now legitimately run 3-5 minutes (8 native pages plus one truncation
retry at a doubled token cap), so the sweep stole the row and the CAS
discarded the worker's real result. Cutoff raised to 10 minutes.

Also: pages_partial_note made period-agnostic (old rows were extracted
from first-pages-only slices, so naming the last page was retroactively
wrong for them), and two stale first-pages-only comments updated.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012xosyW53HUa9JoFiDayhSk

* fix(invoice-inbox): keep the first extraction response when the retry throws

CodeRabbit finding on #2014: a throttled/failed retry call bubbled to the
outer catch before rawText was assigned, discarding a first response whose
text may parse fine despite the truncation flag. The retry is now caught
locally (logged as ai_extraction_retry_failed) and the first result flows on.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012xosyW53HUa9JoFiDayhSk

---------

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

DECISIONS.md

@@ -1323,6 +1323,7 @@ One line per decision: `[YYYY-MM-DD] <decision>: <why>`. Appended by agents and
 [2026-08-28] AR-PDF minus fix uses ASCII hyphen formatting, not font embedding: registering a Unicode TTF for react-pdf would change the whole document's typography and bundle size to fix one glyph; formatPdfKronor keeps built-in Helvetica and sidesteps WinAnsi's missing U+2212.
 [2026-08-28] Same-bank warning limited to observed one-session banks (SEB only): prod shows Handelsbanken tolerates 4 concurrent sessions, and the generic warning made a user abandon a legitimate renewal. Planned sync-death visibility work was dropped: already shipped via #1271 (health probe), #1727 (stale state), #1969 (cron unstarve).
 [2026-08-28] Same-bank warning revised to three tiers after skeptic refutation: hard warn SEB, silent/calm only for verified multi-session banks (Handelsbanken, 4 distinct session_ids observed), legacy hedged warning for unknown banks (fail closed), shared-session siblings exempt (fan-out carries them).
+[2026-08-28] PDF auto-extract page budget: raised to 8 on pdf-native backends and slice now keeps the last page: the 3-page cap was Bedrock-latency legacy (issue #553), but kept a cap at all so 100-page statements do not burn tokens; last page kept because totals/OCR sit there. Truncation retry doubles AI_EXTRACTION_MAX_TOKENS once instead of raising the default: keeps steady-state cost flat.
 [2026-08-28] Per-run salary agent gap: new gnubok_set_run_salary staged tool + v1 PATCH instead of extending gnubok_update_payslip_line: the base-salary value lives on salary_run_employees (draft-gated), not on the payslip line (review-agnostic display copy); overloading the line tool would hide the status semantics and keep the recalc-overwrite trap.
 [2026-08-28] Inbox underlag divergence (#1548) is a separate underlag_status field, not a nulled matched_transaction_journal_entry_id: the book-direct and bulk-book routes 409 on an already-booked transaction, so hiding the verifikat would make the rail re-offer a booking that always fails; the UI keeps divergent items in Att göra with an explanation and a link instead. Anchored-elsewhere conflicts are counted and logged by the daily reconcile cron, never auto-resolved (moving a document between verifikat is a human decision; never-steal is the 2026-08-13 invariant), and not escalated to processing_history from the inline booking path (no dedupe key; it would fire on every booking). The reconcile's 'InboxUnderlagReconciled' event needed a processing_event_types row (FK): the script's old 'InboxUnderlagBackfilled' type was never registered, so its appends had always failed silently.
 [2026-08-28] Inbox underlag reconcile (#1548) bounds link work, not the read: the candidate set (matched, unconsumed) holds permanent residents (samlingsverifikat siblings, anchored-elsewhere items) that never leave it, so a read cap ordered by uuid would revisit the same window nightly and starve the tail. The full scan is four columns per row; maxItems now caps unlinked items linked per run, the rest are counted as deferred; transactions whose items already read anchored are still propagated outside that budget, because only the propagation anchors the transaction's pinned document and stamps settled items out of the scan (idempotent, self-shrinking). A verifikat in a locked/closed period is its own status (unlinked_locked): the period-lock trigger rejects the link every time, so it is neither retried nor promised to the user as automatic. An unreadable document row is reported as 'unknown' on the wire and kept out of the booked bucket, matching the helper's absence-is-never-anchored contract.