Running the stack on built output stopped the idle CPU load, but the
remaining findings from the deployment report were still open.
The Dockerfile is now multi-stage. The prod target runs on node:24-slim as
the "node" user instead of root and carries only its runtime dependencies;
"npm prune --omit=dev" drops tsx, drizzle-kit and typescript. The migration
scripts need drizzle-orm and better-sqlite3, both regular dependencies, so
they keep working. The dev target keeps devDependencies for compose.dev.yaml.
Both compose files get a mem_limit. The server stack also sets memswap_limit
to the same value, so a leak lets the OOM killer replace the container
instead of pushing the host into swap. The Express API is no longer
published on the host: it has no authentication and is reached through the
Next.js rewrite. Development ports are bound to the loopback interface.
Services drop all capabilities and run with no-new-privileges.
SIGTERM and SIGINT now close the HTTP server and the SQLite handle and exit,
instead of only logging the signal and waiting for the container's SIGKILL
ten seconds later. server.closeIdleConnections() releases the keep-alive
sockets the Next.js server holds, which would otherwise keep server.close()
pending for the whole grace period. The compose start commands exec into
node, because otherwise sh stays the foreground process and the signal never
reaches it.
Delete scripts/docker-start.sh rather than fixing its unsupervised
background API: both compose stacks set their own command, so the
single-container entrypoint has no caller left. Also remove drizzle.config.js
(a stray tsc artifact drizzle-kit never reads) and the pre-Next index.html
and styles.css.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Synchronizing a project device pushes its quantity onto every linked
device row, but manualQuantity is not part of the sync snapshot and was
left untouched. A device whose quantity is lower than the row's left the
row at manualQuantity > quantity, violating the snapshot invariant.
The violation surfaced far from its cause: the invariant is only checked
when a full state snapshot is read, and the automatic snapshot runs every
25 revisions. A project could therefore accumulate the broken row
silently and then reject every subsequent command, because the failing
validation rolls back the whole transaction including the revision bump.
Recompute manualQuantity as the synchronized quantity minus the total of
the linked external objects, matching the invariant that
assertCircuitDeviceRowQuantity enforces elsewhere, and reject a
synchronized quantity that falls below that total.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Full-codebase review turned up five real correctness/security bugs and
a dozen smaller inconsistencies; all are fixed here with matching test
coverage:
- BMK uniqueness silently allowed German-umlaut duplicates ("Ä1" vs
"ä1") because the DB's normalized index only folds ASCII case. Added
a shared Unicode-aware pre-check used by every circuit/component
insert and rename path (one of which had no pre-check at all).
- CircuitDeviceRow.simultaneityFactor had no upper bound at the row
level (command model and snapshot/restore schema), unlike every
sibling entity, letting a bad value silently corrupt power totals.
- Grid cell editing silently misread German thousands-separator input
("1.500" parsed as 1.5); "." is now rejected outright with a clear
message instead of guessing.
- The editor's shared command runner (runCommand/applyHistory) had no
re-entrancy guard, so a double click/drop could fire the same
command twice and race a BMK collision or revision conflict. Added a
synchronous ref guard plus isSaving on the buttons that lacked it.
- GET .../next-identifier leaked circuit-numbering state for sections
in other projects (no ownership check, 400 instead of 404). Moved
under /projects/:projectId and scoped it.
Also: added the missing circuits.section_id / circuit_device_rows.
circuit_id indexes (migration 0006), gave FormModal a focus trap /
Escape-to-close / focus restore and rebuilt ProjectSettingsModal on
top of it instead of duplicated markup, removed dead code (3 orphaned
domain model files, an unused persistence helper, a wrapper only used
by its own test), pointed the project page at GET /projects/:id
instead of listing+filtering client-side, closed the gap between the
documented 18 MB CSV limit and the ~17.17 MiB actually enforced, added
missing upper bounds on several free-text fields, filled in nine
missing German labels in the revision timeline, replaced a
key-order-fragile JSON.stringify equality check with a real field
comparison, made an implicit sort-order assumption in three
renumbering helpers explicit, cleared the sidebar's target selection
when it no longer resolves after a tree reload, and fixed
updateGlobalDevice to check-then-write instead of write-then-check.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
