15 of 17 findings from the consolidated audit (3 reviewer agents on
the previously-shipped phase commits). Remaining two are nice-to-have
follow-ups deferred.
Critical (data integrity / security):
- Public berths API: closed-deal junction rows no longer flip a berth
to "Under Offer" - filter on `interests.outcome IS NULL` so won/
lost/cancelled don't pollute public-map status. Both list +
single-mooring routes.
- Recommender heat: cancelled outcomes now count as fall-throughs
(SQL was `LIKE 'lost%'` which silently dropped them, leaving
cancelled-only berths stuck in tier A).
- Filesystem presignDownload returns an absolute URL (origin from
APP_URL) so emailed download links resolve from external mail
clients.
- Magic-byte verification on the presigned-PUT path: both per-berth
PDFs and brochures stream the first 5 bytes via the storage backend
and reject + delete on `%PDF-` mismatch (was only enforced when the
server saw the buffer; presign-PUT was wide open).
- Replay-protection TTL aligned to the token's own expiry (was a
fixed 30 min, but send-out tokens live 24 h). Floor 60 s, ceiling
25 days.
- Brochures unique partial index on (port_id) WHERE is_default=true
+ 0032 migration. Closes the read-then-write race in the create/
update transactions.
Important:
- Recommender SQL: defense-in-depth `i.port_id = $portId` filter on
the aggregates CTE.
- berth-pdf service: per-berth pg_advisory_xact_lock around the
version-number SELECT + insert. Storage key is now UUID-based so
concurrent uploads can't collide on blob paths. Replaces
`nextVersionNumber` with the tx-bound variant.
- berth-pdf apply: rejects with ConflictError when parse_results
contain a mooring-mismatch warning unless the caller passes
`confirmMooringMismatch: true` (force-reconfirm gate was UI-only).
- Send-out body: HTML-escape brochure filename in the download-link
fallback (XSS guard).
- parseDecimalWithUnit rejects negative numbers.
- listClients DISTINCT ON for primary contact resolution: bounds
contact-row count to ~2 per client.
Defensive:
- verifyProxyToken rejects NaN/Infinity expiries via Number.isFinite.
- Replaced sql ANY() with inArray() in interest-berths.
Tests: 1145 -> 1163 passing.
Deferred: bulk-send rate limit (no bulk endpoint today), markdown
italic regex breaking links with asterisks (cosmetic).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Plan §4.4 + §13: pure SQL recommender, no AI. Single CTE chain
(feasible -> aggregates) + JS-side tier classification, fall-through
cooldown filter, heat scoring, and fit ranking. Per-port settings via
system_settings layered over global + DEFAULT_RECOMMENDER_SETTINGS.
Tier ladder (default):
A : no interest history
B : lost-only history (still recommendable + boosted by heat)
C : active interest in early stage (open..eoi_signed)
D : active interest at deposit_10pct or beyond (hidden by default)
Heat (only for tier B):
recency weight 30 full @ <=30 days, decays to 0 @ 365 days
furthest stage weight 40 full when prior reached deposit
interest count weight 15 saturates at 5+
EOI count weight 15 saturates at 3+
Multi-port isolation enforced (§14.10 critical): the SQL filters by
port_id AND the entry-point function rejects cross-port interest
lookups with an explicit error. Fall-through policy supports
immediate_with_heat (default), cooldown, and never_auto_recommend.
15 unit tests covering tier classification, heat saturation, weight
tuning, zero-weight guard. Smoke-tested end-to-end via
scripts/dev-recommender-smoke.ts.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
