fix(ui): restore the trades card on the operator overview

[argszero]

Summary

Restore the trades card on the operator overview.

GET /api/ops/runtime has returned total_txs — a global COUNT(*) FROM transactions — since 85982e8 (PR #80, v0.3.3). The live operator view never rendered it. The mock branch did:

stat(T("ops.stats.trades"), T("cnt.trades", { n: txs.length }), T("ops.stats.trades.sub"))

and 89963f3 ("refactor(ui): zero mock data in logged-in views…", v1.22) deleted that line without re-pointing the live branch at rt.total_txs, which the same response object already carried.

So this is a regression being repaired, not a new feature. It is also not prototype parity: the prototype has four ops cards and no trades card (交易 occurs 0 times in docs/prototype/aitokenpool-console.html), while the live view has nine cards by design. Nothing is trimmed toward the prototype.

The card's two i18n keys (ops.stats.trades, ops.stats.trades.sub) already existed in both packs — that is why this change adds no keys, and the language-pack gate's key counts stay at 786/786.

Related Issue

No issue exists for this (found by a repo-wide orphan-key and unconsumed-response-field sweep). Filed up front instead.

Changes

• ui/js/app.js: the live card list renders total_txs via the pre-existing keys, placed directly after the keys card — the position the mock branch had.
• src/routes/mod.rs: new route test ops_runtime_total_txs_is_global_and_all_types.
• src/i18n_pack.rs: the maintained call-site controls move 535 → 538 and 429 → 431 (3 new T() calls; cnt.trades was already distinct via the dashboard). Key counts deliberately unchanged at 786 / 786 — the tripwire that proves no key was added.
• ui/index.html: cache-bust ?v=20260912-1 → ?v=20260912-2 (5 refs).
• ui/README.md: documents the restored card and why it is not a prototype alignment.
• No config/data-structure change, so config.example.toml needs no sync.

Tests

• cargo test — 165 passed / 0 failed (164 + 1 new)
• cargo fmt --check — clean
• cargo clippy --all-targets -- -D warnings — 0 warnings
• New unit test added (see below)

Why the field had no test, and why that mattered: grep total_txs src/routes/ops.rs found it only at the compute line and the serialize line — zero assertions in either direction. That absence is precisely why a load-bearing field could go unused for 23 commits. The new test pins both facts: that it is returned, and that its scope is the whole table across all types. The three rows it creates span two users (admin's topup via the real credits endpoint + demo's consume/earn), so the expected 3 differs from both a consume-only count (1) and a per-user max (2) — either degradation is falsifiable, and both are asserted as explicit counter-checks.

Teeth verified by injection: degrading the query to WHERE type = 'consume' reddens the new test with total_txs: 1 (expected 3); restored byte-identically afterwards (md5-verified).

A/B runtime proof (24/24): both legs generated from the real #ops-stats expression and the real fmtUptime, with the real pre-change source (git show HEAD:ui/js/app.js) as the before leg — not a perturbation of the after source. BEFORE 8 cards, no trades label; AFTER 9 cards, value rendered from the response (12345 → 12345 笔 / 12345), subtitle 累计全部类型 / all types, positioned right after the keys card. Changing total_txs to 777 moves the card, proving the value is not a constant; the mock row count (8) is asserted absent, keeping the zero-mock rule intact.

Checklist

• Branch naming follows the convention (fix/)
• Commit message uses Conventional Commits
• Single responsibility, minimal change