feat(agent): resizable, undockable assistant panel (#1467)

* feat(agent): resizable, undockable assistant panel

User report: the assistant chat sheet sometimes covers the page content
the user is asking about, with no way to resize or move it.

- Docked mode is now drag-resizable from its left edge (380-800px,
  clamped so the page keeps a 480px readable column) and the page
  reflows beside it via the existing --agent-dock-w reservation.
- Expanded (focus) mode reserves page margin like the compact dock
  instead of overlaying up to 1100px of the page.
- New undock toggle turns the sheet into a floating window that can be
  dragged by its header and resized from edges/corners, clamped so the
  header always stays reachable. Desktop only; mobile keeps the
  full-screen sheet.
- Geometry (mode, dock width, float rect) persists per user in
  user_preferences.ui_state.agent_panel, server-seeded to avoid a
  first-paint jump; the ui-state API schema gains a strict agent_panel
  key with nested merge.
- Pure clamp/resize math lives in lib/agent-panel/geometry with unit
  tests; drag frames write styles imperatively and commit one
  preference update on release.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(agent): address review findings on panel drag, a11y, and persistence

CodeRabbit round 1, all six findings fixed:

- Bind drag listeners to window (plus lostpointercapture) so a failed
  pointer capture or mid-drag unmount can never leave the transition
  suppression and data-agent-resizing stuck for the session.
- Keyboard resize now steps from the visible width (expandedW in focus
  mode) instead of jumping to the persisted dock width.
- The width handle exposes window-splitter semantics: aria-valuenow,
  aria-valuemin, aria-valuemax.
- --nav-w is read reactively via a MutationObserver on #dash-shell
  instead of computed-style reads in the render body and per drag frame.
- The ui-state POST in updatePanelPrefs gets a 300ms trailing debounce
  (state stays immediate) so key auto-repeat cannot produce one
  read-merge-write per repeat; pending write flushes on unmount.
- globals.css keeps one :root token block; the agent-resizing rule moved
  below it.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(agent): filter drag events by pointer id, clear fired debounce timer

CodeRabbit round 2, both findings fixed:

- Window-level drag listeners now ignore events from pointers other than
  the initiating one, so a second touch or pen cannot move the panel or
  end the first pointer's drag.
- The persist debounce timer ref is nulled when the timer fires, so the
  unmount flush only writes genuinely pending values instead of
  replaying an already-persisted (possibly stale) geometry.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

---------

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

DECISIONS.md

@@ -833,6 +833,7 @@ One line per decision: `[YYYY-MM-DD] <decision>: <why>`. Appended by agents and
 [2026-07-28] Transaction method (structured payment rail): the trailing channel phrase ("Överföring via internet", "Kortköp/uttag") is stripped from transactions.description at INGEST and by a one-shot BACKFILL, not merely hidden at render: description is the mutable working title, original_description keeps the full bank string, and every dedup surface (external_id: date+öre only; content bridge: prefix-containment over original_description ?? description, and a trailing strip leaves a prefix) is provably unaffected. transaction_method is text + CHECK (repo convention, no PG enums) beside verbatim bank_transaction_code / proprietary_bank_transaction_code evidence columns per data_quality_master Appendix B Layer-A; the dead `enrichment` jsonb was NOT reused (the Gokind lesson: opaque blobs with no readers die). mapping-engine now also matches original_description so user rules written against the full bank text keep firing on stripped rows.
 [2026-07-29] Transaction-method backfill scope: classification and title-stripping are FEED-ROW concepts (import_source present, not manual/mcp), enforced identically at ingest and in the 20260808090100 backfill, plus an adjective guard so "Egen insättning"/"Eget uttag"/"Intern överföring" keep their full titles even on feed rows (the phrase IS the meaning after a possessive/scope adjective). Chosen over vocabulary tweaks because the failure mode for unknown bank phrasings must be "row unchanged", and user-authored titles must never be rewritten by a channel vocabulary. A read-only prod dry-run script exists for coverage measurement but prod reads were left to the founder (permission-gated).
 [2026-08-08] Compliance-bot finding on the transaction_method backfill (booked rows' titles rewritten without a rattelse trail) triaged as satisfied-by-design, not a blocker: BFL 5 kap 5 attaches to bokforingsposter, and the backfill touches no journal table; the verifikat description is snapshotted into journal_entries at commit and SIE #VER export reads journal_entries only (both verified in code, no report reads transactions.description lazily); the bank original is preserved byte-identical in original_description by the same UPDATE (enforced since 80ef1ee0, and prod has 0/25,566 feed rows lacking it). The stricter TRANSACTION_TITLE_LOCKED gate on booked rows blocks arbitrary user free-text renames, a different mutation class from a deterministic trailing-vocabulary strip that skips user-edited titles and keeps the original adjacent. Period-lock triggers sit on the journal tables and fiscal periods, not on transactions; the pg-upgrade CI run applied the backfill against seeded booked rows with all enforcement triggers active.
+[2026-08-08] Assistant panel undock/resize hardcodes Swedish UI strings (title/aria on the new handle and dock/undock toggle) instead of next-intl: every existing string in components/agent (AgentSheet, AgentChat, session list) is hardcoded Swedish with zero next-intl usage, so two translated strings in an otherwise Swedish-only surface would be inconsistent both ways; migrating the whole agent surface to next-intl is its own change. Geometry persists in user_preferences.ui_state.agent_panel (design rule 9, never localStorage); expanded focus mode now reserves page margin like the compact dock instead of overlaying (the reported bug), clamped so the page keeps a 480px readable column; @use-gesture/react stayed unused in favor of a ~30-line pointer-capture helper (three call sites need only delta tracking).
 [2026-08-08] Receipts checklist step (PR #1461) done-signal = any invoice_inbox_items row for the company, not a document_attachments count: document_attachments also archives outbound invoice PDFs, ROT/RUT payout files and Bolagsverket submissions, so its count would mark the step done when the user merely sends their first invoice; invoice_inbox_items is written by exactly the three intake channels (email/WhatsApp/upload) and carries a company_id index. Same PR: vat_registered with moms_period=null renders an explicit "välj momsperiod"-link in the checklist instead of silence, because that state makes the deadline engine generate zero VAT deadlines without any error (all three moms configs require a concrete period).
 [2026-08-08] Per-voucher RC basis gaps (findRcBasisGaps) downgrade from filing-blocking ERROR to WARNING only under per-rate evidence: the 44xx/45xx basis accounts grouped by momssats must match ruta30/0.25, ruta31/0.12, ruta32/0.06 two-sided within 0.5 kr, all rate boxes non-negative, and no RC_OUTPUT_MISSING present. A first cross-rate-sum predicate was refuted by /skeptic (wrong-rate fiktiv moms reached parity and unblocked a 7 800 kr under-declaration; a negative rate box made the predicate vacuous), so the certificate is per-rate, which rutor alone cannot express (rutor 20-24 are partitioned by purchase type, not rate); evidence therefore flows from the account totals. Why downgrade at all: a moms-only rattelseverifikat carries fiktiv moms whose basbelopp lives in another (often reversed) verifikat, and no voucher arrangement satisfies both the per-voucher scan and the aggregate identity in that state, so the ERROR was an unfixable dead end (Orto Engineering 3DJake case 2026-08; support vouchers A169/A175/A177 joined the blocklist they were meant to clear). Data side repaired separately with voucher A177 restoring bank parity and the basis/moms identity exactly.
 [2026-08-08] Fenced-JSON fix uses brace-slice, not fence-regex: also rescues preamble/postamble prose around the object, and degrades to the existing empty-result path when no braces exist.