feat(ui): let the transactions trend card switch between four metrics

[argszero]

What this is

The transactions page trend card #tx-trend unconditionally drew two bars per
bucket (spend + income) and offered no control at all. The design prototype
(docs/prototype/aitokenpool-console.html) draws the same two bars, so this is a
new feature request, not a control that was lost when the card was built.

The card gets a .tabs segment — #tx-trend-modes, the shape already used by
#tx-tabs — with four modes:

mode	rendering
total points (default)	line, cumulative net change over the window
spend	bars, spend series only
income	bars, income series only
spend + income	bars, both series

No backend change, and no extra request

/api/transactions/trend (src/routes/wallet.rs) already applies the range, the
type tab and the column filters, and already returns net, income and
expense per bucket. All four modes are therefore derived from the payload the
card already has.

Consequently the mode is not part of the request: it does not enter
txQuerySig() and the mode buttons do not call reloadTransactions().
Folding it into the reload trigger would issue a fresh request on every click and
make the four modes stop sharing one payload.

The mode that is easiest to get wrong

"Total points" is the cumulative net change inside the window: the running sum
of net from the earliest bucket, starting at 0, and it can be negative. Under
"all time" with no column filter its last point equals the real current balance —
but it is not an absolute balance level, and with a column filter applied an
absolute level is not even defined ("the balance of model=deepseek" does not
mean anything).

So that mode swaps the card title and subtitle for tx.trend.card.net and
tx.trend.net.sub, which state the definition outright, and draws a line
(sparkline()) rather than bars — bars would relativise the level and make the
balance look like it keeps resetting to zero. The y domain is taken from the
signed range and keeps the zero baseline inside it. The other three modes keep
the existing bar rendering and normalisation.

Self-consistency of the four modes

• the normalisation max is taken from the series actually displayed (so
  looking at income alone is not flattened by a spend peak);
• the title tooltip and the legend name only the displayed series — a
  hidden series does not linger in the legend or the tooltips;
• the three empty states are unchanged (not loaded / load failed / no records in
  window).

Where the reset lives

Each entry to the transactions view returns to the default mode. The reset is at
the view entry in switchView(), not in renderView() or
renderTxTrend(): those two are also reached from the atp:langchange
handler, so hanging the reset on the render chain would silently reset the user's
chosen mode on a language switch. Switching is not persisted across sessions.

i18n

Four keys added to both language packs (tx.trend.mode.net,
tx.trend.mode.both, tx.trend.card.net, tx.trend.net.sub); the two existing
metric keys (tx.trend.metric.expense, tx.trend.metric.income) are reused on
the buttons. No existing value was changed. Cache-bust query bumped to
20260924-1 on css/style.css, js/i18n.js and js/app.js.

Scope

Only this one card on the transactions page. The dashboard card
(#dash-trend / renderDashTrend()) is deliberately untouched.

Evidence

Three instruments, all run against the tree in this PR; the pre-fix tree is
materialised from the parent commit (git archive 1f1bd68 ui).

• logic probe (r109_logic_probe.js) — legs=9 failed=0; the 6 synthetic
  mutations each fail exactly the leg they target.
• render probe (r109_render_probe.js, jsdom, applies the edit sheet itself):
  pre-fix tree exactly 15 of 15 legs red; fixed tree legs=16 failed=0.
  Its edit output is byte-identical to the landed files —
  app.js 908aeb05… → 12788e38…, index.html 6b8365b9… → 8b14c8d3…,
  i18n.js 3cbd93d9… → e61c006e…, css/style.css e5dd3518… → be072471….
• end-to-end probe (r109_e2e_probe.js, real index.html + the four real
  scripts + a real login + real click()): pre-fix tree 7 of 7 legs red;
  fixed tree legs=7 misdeclared=[]. It also checks the two edit tables
  agree (L0) and, separately, that clicking a mode does not reach the list at
  all — no request issued and the pager stays on page 2 (L4) — which a
  request count alone would miss, since the wrong wiring
  (reloadTransactions()) sends no request but still throws the user back to
  page 1.
• cargo test: 373 passed / 0 failed (unchanged from the parent — this PR
  adds no Rust gate). cargo fmt --check clean.
• cargo clippy --all-targets -- -D warnings: rc=101 from a pre-existing
  collapsible_match at src/protocol.rs:662, which a pristine parent-commit
  tree reports identically. Not part of this diff; left untouched.