Replaces the legacy text-only expense PDF (was just dumping rows into a
single pdfme text field — no images, no pagination) with a proper
streaming export modelled on the legacy Nuxt client-portal but
re-architected for memory safety. The legacy implementation OOM'd on
hundreds of receipts because it:
- buffered every receipt image into memory simultaneously
- accumulated PDF chunks into an array, concat'd at end
- base64-encoded the whole PDF into a JSON response (3x peak memory)
- had no image downscaling
The new design:
- `streamExpensePdf()` (src/lib/services/expense-pdf.service.ts):
pdfkit pipes bytes directly to the HTTP response (no Buffer
accumulation). Receipts are processed serially so peak heap is one
image at a time. Sharp downscales any receipt > 500 KB or > 1500 px
to JPEG q80 — typical 8 MB phone photo collapses to ~250 KB. For a
500-receipt export, peak RSS stays under ~100 MB; legacy needed >2
GB for the same input.
- Pages: cover summary box (count, totals, currency equiv, optional
processing fee), grouped expense table (groupBy=none|payer|category|
date), one-page-per-receipt with header (establishment, amount,
date, payer, category, file name) and full-bleed image.
- Storage backend abstraction — receipts stream from
`getStorageBackend().get(storageKey)`, works on MinIO/S3/filesystem.
- Route: POST /api/v1/expenses/export/pdf streams binary
application/pdf with cache-control:no-store. Validator caps
expenseIds at 1000 to prevent runaway loops.
Receipt-less expense flow (per user request):
- Schema: 0033 migration adds `expenses.no_receipt_acknowledged`
boolean (default false).
- Validator: createExpenseSchema requires either receiptFileIds OR
noReceiptAcknowledged=true; the .refine() error message tells the
rep exactly what to do. updateExpenseSchema is partial and skips
the rule (existing rows can be edited without re-acknowledging).
- PDF: receiptless expenses get an inline red "(no receipt)" tag in
the establishment cell + a red footer warning in the summary box
showing the count and at-risk amount.
- The legacy parent-company reimbursement queue may refuse to pay
receiptless expenses, so the warning is load-bearing for ops.
Audit-3 fixes piggy-backed:
- 🔴 Tesseract OCR runtime now races a 30s timeout (CPU-bomb DoS
protection — a crafted PDF rasterizing to high-res noise could
pin the worker indefinitely).
- 🟠 brochures.service.ts:listBrochures dropped a wasted query (the
legacy single-brochure fast-path was discarding its result on the
multi-brochure branch).
- 🟠 berth-pdf.service.ts:listBerthPdfVersions now Promise.all's the
presignDownload calls instead of awaiting each in a for-loop —
20-version berths went from 20× round-trip to 1×.
- 🟡 public berths route no longer logs the full `row` object on
enum drift (was dumping price + amenity columns into ops logs).
- 🟡 dropped the dead `void sql` import from public berths route.
Tests still 1163/1163. tsc clean.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three independent strengthenings of the sales spine that the prior coherence
sweep made it possible to do cleanly.
1. EOI queue page
- Sidebar entry under Documents → "EOI queue".
- Route /[port]/documents/eoi renders DocumentsHub with the existing
eoi_queue tab pre-selected (filters in-flight EOIs only).
- .gitignore: tightened root-only `eoi/` ignore so the documents/eoi
route is no longer silently excluded.
2. Invoice ↔ deposit link
- invoices.interestId (FK, ON DELETE SET NULL) + invoices.kind
('general' | 'deposit'). Indexed on (port_id, interest_id).
- createInvoiceSchema requires interestId when kind === 'deposit';
the service validates the linked interest belongs to the same port
before insert.
- recordPayment auto-advances pipelineStage to deposit_10pct (via
advanceStageIfBehind) when a paid invoice is kind=deposit and has
an interestId. No-op if the interest is already further along.
- "Create deposit invoice" link added to the Deposit milestone on the
interest detail. Links to /invoices/new?interestId=…&kind=deposit;
the form prefills the billing entity from the linked interest's
client and shows a context banner.
3. Won / lost terminal outcomes
- interests.outcome ('won' | 'lost_other_marina' | 'lost_unqualified'
| 'lost_no_response' | 'cancelled') + outcomeReason text +
outcomeAt timestamp. Indexed on (port_id, outcome).
- setInterestOutcome / clearInterestOutcome services + POST/DELETE
/api/v1/interests/:id/outcome endpoints (gated by change_stage
permission). Setting an outcome moves the interest to `completed`
in the same write; clearing reopens to `in_communication` (or a
caller-specified stage).
- Mark Won / Mark Lost icon buttons on the interest detail header,
plus an outcome badge that replaces the stage pill once a terminal
outcome is set, plus a Reopen button.
- Funnel + dashboard math updated to exclude lost/cancelled outcomes
from active calculations (KPIs.activeInterests, pipelineValueUsd,
getPipelineCounts, computePipelineFunnel, getRevenueForecast).
The funnel now also returns a `lost` summary so callers can
surface leakage without polluting conversion percentages.
Schema changes shipped via 0019_lazy_vampiro.sql; applied to dev DB
manually via psql because drizzle-kit push hits a pre-existing zod
parsing issue on the companies index. Dev server may need a restart
to flush prepared-statement caches.
tsc clean. vitest 832/832 pass. ESLint clean on every file touched.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
PR1 of Phase B per docs/superpowers/specs/2026-04-28-phase-b-insights-alerts-design.md.
Lays the foundation that PRs 2-10 will fill in with behaviour.
Schema (migration 0014):
- alerts table with rule-engine fields (rule_id, severity, link,
entity_type/id, fingerprint, fired/dismissed/acknowledged/resolved
timestamps, jsonb metadata). Partial-unique fingerprint index keeps
one open row per (port, rule, entity); separate indexes power
severity-filtered and time-ordered queries.
- analytics_snapshots (port_id, metric_id) -> jsonb cache + computedAt
for the 15-min recurring refresh.
- expenses: duplicate_of self-FK, dedup_scanned_at, ocr_status/raw/
confidence; partial index on (port, vendor, amount, date) where
duplicate_of IS NULL drives the dedup heuristic.
- audit_logs.search_text: GENERATED ALWAYS tsvector over
action+entity_type+entity_id+user_id, GIN-indexed (drizzle can't
model GENERATED ALWAYS in TS yet, so the migration appends manual
ALTER + the GIN index).
Service skeletons in src/lib/services/:
- alerts.service.ts: fingerprintFor, reconcileAlertsForPort (upsert +
auto-resolve), dismiss, acknowledge, listAlertsForPort.
- alert-rules.ts: RULE_REGISTRY of 10 rule evaluators (currently no-op);
PR2 fills in the bodies.
- analytics.service.ts: readSnapshot/writeSnapshot with 15-min TTL +
no-op compute* stubs for the four chart series; PR3 fills behavior.
- expense-dedup.service.ts: scanForDuplicates + markBestDuplicate
using the partial dedup index. PR8 wires the BullMQ trigger.
- expense-ocr.service.ts: OcrResult/OcrLineItem types + ocrReceipt
stub. PR9 wires Claude Vision (Haiku 4.5 + ephemeral system-prompt
cache).
- audit-search.service.ts: tsvector @@ plainto_tsquery + cursor
pagination on (createdAt, id). PR10 wires the admin UI.
tsc clean, lint clean, vitest 675/675 (one unrelated AES random-output
flake passes solo).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
