Two reviewer agents did a second-pass deep audit of the 21-commit
refactor. Eight findings; four fixed here (one was deferred with a
schema comment, three were 🟡 nice-to-haves left for follow-up).
Integration regressions (🟠 high):
- Outbound webhook `interest.berth_linked` now fires from the new
junction-add handler. Was emitting a socket-only event, leaving
external integrations silent post-refactor.
- Two new webhook events `interest.berth_unlinked` and
`interest.berth_link_updated` added to WEBHOOK_EVENTS +
INTERNAL_TO_WEBHOOK_MAP. PATCH and DELETE handlers now dispatch them
alongside the existing socket emits — lifecycle parity restored.
- BerthInterestPulse adds useRealtimeInvalidation for berth-link
events. The query key was berth-scoped while the linked-berths
dialog invalidates interest-scoped keys (no prefix match), so the
pulse went stale. Bridges via the realtime hook now.
Recommender semantic fix (🟠 medium-high):
- aggregates CTE: active_interest_count now filters on
`ib.is_specific_interest = true`, matching the public-map "Under
Offer" derivation. EOI-bundle-only links no longer demote a berth
to Tier C for other reps. Smoke test confirms previously-all-Tier-C
results now correctly classify as Tier A.
- Same CTE: `total_interest_count` uses COUNT(ib.berth_id) instead of
COUNT(*) so a berth with no junction rows reports 0 (not 1 from
the LEFT JOIN's NULL-right-side row). Prevents heat over-counting.
Data integrity (🟠):
- AcroForm tier rejects negative numerics in coerceFieldValue (was
letting through `length_ft="-50"` which would poison the
recommender feasibility filter on apply).
- FilesystemBackend.resolveHmacSecret throws in production when
storage_proxy_hmac_secret_encrypted is null. Dev still derives from
BETTER_AUTH_SECRET for ergonomics; prod must explicitly configure.
- Documented the circular FK between berths.current_pdf_version_id
and berth_pdf_versions.id. Drizzle's `.references()` can't express
the cycle so the schema column is plain text + a comment; the FK
is authoritatively maintained by migration 0030.
Tests still 1163/1163. tsc clean.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Phase 6b of the berth-recommender refactor (see
docs/berth-recommender-and-pdf-plan.md §3.2, §3.3, §4.7b, §11.1, §14.6).
Builds on the Phase 6a pluggable storage backend (commit 83693dd) — every
file write goes through `getStorageBackend()`; no direct minio imports.
Schema (migration 0030_berth_pdf_versions):
- new table `berth_pdf_versions` with monotonic `version_number` per
berth, `storage_key` (renamed convention from §4.7a), sha256, size,
`download_url_expires_at` cache slot for §11.1 signed-URL throttling,
and `parse_results` jsonb for the audit trail.
- new column `berths.current_pdf_version_id` (deferred from Phase 0)
with FK to `berth_pdf_versions(id)` ON DELETE SET NULL.
- relations + types exported from `schema/berths.ts`.
3-tier reverse parser (`lib/services/berth-pdf-parser.ts`):
1. AcroForm via pdf-lib — pulls named fields (`length_ft`,
`mooring_number`, etc.) at confidence 1. Sample PDF has 0 such
fields, so this is defensive coverage for future templates.
2. OCR via Tesseract.js — positional/regex heuristics keyed off the
§9.2 layout (Length/Width/Water Depth as `<imperial> / <metric>`,
`WEEK HIGH / LOW`, `CONFIRMED THROUGH UNTIL <date>`, etc.). Returns
per-field confidence + global mean; flags imperial-vs-metric drift
>1% in `warnings`.
3. AI fallback — gated via `getResolvedOcrConfig()` (existing
openai/claude provider). Surfaced from the diff dialog only when
`shouldOfferAiTier()` returns true (mean OCR confidence below
0.55 threshold), so OPENAI_API_KEY isn't burned on every upload.
Service layer (`lib/services/berth-pdf.service.ts`):
- `uploadBerthPdf()` — magic-byte check, size cap, version-number
bump + current pointer in one transaction.
- `reconcilePdfWithBerth()` — auto-applies fields where CRM is null;
flags conflicts when CRM and PDF disagree; tolerates ±1% on numeric
columns; warns on mooring-number-in-PDF mismatch (§14.6).
- `applyParseResults()` — hard allowlist of writable columns;
stamps `appliedFields` onto `parse_results` for audit.
- `rollbackToVersion()` — pointer flip only, never re-parses (§14.6).
- `listBerthPdfVersions()` — version list with 15-min signed URLs.
- `getMaxUploadMb()` — port-override → global → default 15 lookup
on `system_settings.berth_pdf_max_upload_mb`.
§14.6 critical mitigations:
- Magic-byte check (`%PDF-`) on every upload; mismatch deletes the
storage object and rejects the request.
- Size cap from `system_settings.berth_pdf_max_upload_mb` (default
15 MB); enforced in the upload-url presign AND server-side.
- 0-byte uploads rejected.
- Mooring-number mismatch surfaces as a `warnings[]` entry on the
reconcile result so the rep sees it in the diff dialog.
- Imperial vs metric ±1% tolerance in both the parser warnings and
the reconcile equality check.
- Path traversal already blocked at the storage layer (Phase 6a).
API + UI:
- `POST /api/v1/berths/[id]/pdf-upload-url` — presigned URL (S3) or
HMAC-signed proxy URL (filesystem) sized to the per-port cap.
- `POST /api/v1/berths/[id]/pdf-versions` — verifies the upload via
`backend.head()`, writes the row, bumps `current_pdf_version_id`.
- `GET /api/v1/berths/[id]/pdf-versions` — version list + signed URLs.
- `POST /api/v1/berths/[id]/pdf-versions/[versionId]/rollback`.
- `POST /api/v1/berths/[id]/pdf-versions/parse-results/apply` —
rep-confirmed diff payload.
- New "Documents" tab on the berth detail page (`berth-tabs.tsx`)
with current-PDF panel, version history, Replace PDF button, and
`<PdfReconcileDialog>` for the auto-applied + conflicts UX.
System settings:
- `berth_pdf_max_upload_mb` (default 15) — caps presigned-upload size
+ server-side validation. Resolved port-override → global → default.
Tests:
- `tests/unit/services/berth-pdf-parser.test.ts` — magic bytes,
feet-inches, human dates, full §9.2-shaped OCR text → 18 fields,
drift warning, AI-tier gate.
- `tests/unit/services/berth-pdf-acroform.test.ts` — synthetic
pdf-lib AcroForm round-trip.
- `tests/integration/berth-pdf-versions.test.ts` — upload, version-
number bump, magic-byte rejection, reconcile auto-applied vs
conflicts vs ±1% tolerance, mooring-number warning,
applyParseResults allowlist enforcement, rollback semantics.
Acceptance: `pnpm exec tsc --noEmit` clean, `pnpm exec vitest run`
green at 1103/1103.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Matt Ciaccio