Thoughtful software, built all the way through.

Under production traffic the app hit intermittent 500s and pages that loaded forever.

PostgreSQL connection-pool exhaustion against Supabase's transaction pooler (a hard 200-connection cap): the pooled postgres.js client was re-created per request instead of cached across warm serverless invocations, and postgres.js has no built-in wait-timeout, so a saturated pool made server components hang rather than fail.

Cache the pooled client across invocations, tune pool size and idle_timeout / connect_timeout, wrap hot-path DB calls in an application-level timeout, and add Next.js error boundaries.

Failures now surface as errors with a retry UI instead of infinite spinners, and the pool no longer exhausts under load.

Naive strike-selection could pick a strike that doesn't exist in the real option chain, or a stale contract with no active market.

Live option-chain data isn't uniformly spaced, some underlyings mix legacy corporate-action-adjusted strikes with the current grid at irregular intervals, so 'round to the nearest step' arithmetic breaks.

Select strikes from the broker's actual listed-strike data rather than assumed spacing, with a bounded, liveness-aware search that prefers a genuinely tradeable nearby strike over a numerically-closer illiquid one.

Falls back to an explicit 'no viable contract' state when nothing nearby is tradeable, so a bad chain is never silently misread as a good one.

On real bank PDFs, reprinted column headers, brought/carried-forward markers and a footer disclaimer's phone numbers were leaking into transactions, and one footer line became a phantom transaction that broke the running-balance chain.

PDF.js returns loosely positioned x/y text fragments, not tables, so rows and columns have to be reconstructed from coordinates, and page-break reprints and footer text sit in the same coordinate space as genuine rows.

Detect and skip header/footer/balance-forward rows, strip leaked ledger markers while preserving genuine leading-digit references, and reconcile each row's amount against the running balance to recover debit/credit direction.

Verified against a real five-page statement: four phantom rows removed and all balance-mismatch warnings cleared.

After a service-worker restart, Panic Mode could invert or screen-share protection could leave tabs permanently blurred.

MV3 service workers are unloaded and restarted at any time, silently resetting in-memory module state like 'is Panic currently on?'.

Move ephemeral cross-restart state into chrome.storage.session, make every disable path DOM-authoritative (always query and clear tracked elements), and add a monotonic generation counter to invalidate superseded async work.

A desynced flag can never leave a page stuck blurred, and racing profile apply/clear operations no longer clobber each other.

The popup could show 'no trackers' on a tab open for a while, even though the badge still showed a stale count.

MV3 service workers terminate after ~30 seconds of inactivity, wiping the in-memory Map that held per-tab detection state.

Serialize the per-tab set of tracker IDs into chrome.storage.session and rehydrate on read; and to make 'Stop' instant on already-running trackers, inject a MAIN-world script that swaps each tracker's live API functions for no-op proxies.

Detection state survives worker restarts within the session, and blocking takes effect immediately without a page reload.