fix(usd): review findings before the trade-time USD read side ships - #137

From a strict review of af4e6e5..133fabd (#133, #136):

- Community pairs are never stored (usdLedger.capturable). A pair priced from an
  outside pool can be pushed for a minute, and a captured price stayed in the
  totals for good. A suspended or delisted pair must show no USD, which a stored
  value defeated. Their trades keep usd NULL and count at the pair's price now,
  which follows its listing status. The backfill (apply) clears pair values
  captured since #133 and drops their ledger rows.
- A launch row whose volume is partly unpriced while its quote has no price
  right now shows its quote amount (launchVolumeUsd: null) instead of a smaller
  dollar figure that looked complete.
- The receipt path (applyLaunchTx, the user is waiting) no longer captures live
  prices: a capture could wait on Coinbase and Chainlink. Those trades take the
  hour's price the poller captured, or the heal pass prices them.
- Heal pass (usdLedger.healUnpricedUsd), every 5 minutes per chain after a sync:
  - prices the last 3 days' trades and quote-side fees with no USD whose hour now
    has a price, and adds them to the launch totals in the same statement
  - skips pairs and re-tests usd IS NULL inside the UPDATE, so two machines
    change each row and each total once
- Backfill apply is ONE transaction under the rebuild-launch-totals locks: pair
  cleanup, pricing and the totals recompute commit together. No reader sees rows
  priced but not yet summed, and a failure changes nothing. The dry run counts
  pairs as unpriced and stays write-free.
- The lists probe one launch at a time (LATERAL u on the new partial index
  bb_launch_swaps_usd_missing_token_idx) instead of aggregating every unpriced
  trade per request. The totals' unpriced-fee read gets
  bb_launch_fee_events_usd_missing_idx.
- scripts/rebuild-launch-totals.sql recomputes the USD totals too and reports
  USD drift before and after.
- Wording: the tooltip and the partial note describe what has no recorded value.
  The backfill retries network errors and timeouts, not only 429/5xx.

Deferred, not risky: the pairs pages and the token fee panel still use today's
price; the volume sort mixes units during an ETH price outage (pre-existing);
per-hour source labels; poolMaps / quoteUsdNow duplication.

Verified on a throwaway Postgres 16 with the real migration (3 partial indexes
built):
- two concurrent heal passes priced the missed trade and fee once (AETH $2,000 →
  $6,200, not $10,400; fee $21); a rerun and a throttled run changed nothing; the
  pair stayed NULL
- the dry run reported pairs as unpriced and changed nothing
- apply (1 BEGIN / 1 COMMIT) cleared the old pair value and its ledger rows and
  left AETH at $6,200
- the repair script ran with 0 USD drift
- the real reads: AETH $6,200, the pair row null (quote amount), sort ok, totals
  $6,200 and partial
tsc and lint clean; npm test 1410 (1399 pass, 11 skipped, 0 fail).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>