fix(ui): label a market row's availability from the row, not from a missing key count

[argszero]

Summary

On the guest marketplace every row's availability cell reads "No key", while on the same row the availability dot is green and the 使用 / Use button is enabled — two pieces of copy on one row contradicting each other.

A market row's availability has exactly one fact: avail. It is what the dot (.dot), the disabled state of the 使用 button, the "Available only" filter and the expanded detail's "当前可用 / 当前繁忙" all render. keys is not a second fact — it only exists when a count is known: the logged-in path fills it from /api/models → available_keys via modelsToView(), and the guest fallback table (data.js > MARKET) deliberately carries no count (rant 2026-08-19T15:54:06: multi / success were removed as fabricated data). The availability pill, however, read only keys, so m.keys || 0 was 0 and every guest row fell into the "no key" branch.

Measured in jsdom with the real ui/index.html + the four real scripts (only fetch stubbed):

• guest leg — 7/7 rows read "No key"; on 6 of them the dot is live and the button is enabled
• logged-in leg (positive control) — the three count states (keys=3/1/0) render correctly, so the pill itself is not broken; what is missing is the field it reads on guest rows

Related Issue

Changes

• ui/js/app.js: availPill() keeps the count states (>=2 multi, ==1 tight) and falls back to the row's own avail when no count is present, so the cell can no longer contradict the rest of the row.
• ui/js/i18n.js: new key mk.avail.on (zh 「可用」 / en "Available") — the honest label when only "usable" is known. mk.avail.multi / mk.avail.tight / mk.avail.none unchanged.
• ui/index.html: cache-bust i18n.js / app.js.
• src/catalog_gate.rs: new gate market_availability_pill_agrees_with_the_row_it_renders pinning both halves — the consumer must consult m.avail (with the fallback table's boolean avail as the positive control that the field name is real), and MARKET rows must not carry a hand-written keys count (that would put fabricated operational data in front of guests). js_function_body_stops_at_the_right_place proves the extractor stops at the right place and that the assertion has teeth via a synthetic input.
• src/i18n_pack.rs: key-count constants follow the +1 key (the extractor's own output was used for calibration, not hand arithmetic).
• ui/README.md: new convention section + key-count sentence kept in sync.

Not the fix: adding hand-written counts to MARKET. That is the same thing v1.22 cleaned up (multi / success); if guests are to see a real count it has to come from /api/models.

Tests

• cargo test — 262 passed / 0 failed (was 260; +2 new gate tests)
• cargo fmt --check — clean
• cargo clippy — only the pre-existing warning at src/protocol.rs:662
• New unit tests added (the gate above)

A/B evidence (tmp/c2128_*, all UI mutations injected from tmp/ through env vars — the work tree is never rewritten):

leg	red set	note
v0_orig (pinned ecc654c app.js/i18n.js)	{B1, B2, B3}	the axis
m_swap (pill reads only m.avail, count states dropped)	{C2}	kills the logged-in positive control ⇒ the count states must stay
m_fabricate (competing fix: hand-written keys in MARKET)	{}	honestly recorded: the probe alone does not pin the direction — the Rust gate does (its MARKET-must-not-carry-keys half goes red)
live	{}	11/11

In-place gate legs (byte-identical restore verified by md5): m_nofallback → red at the consumer assertion; m_fabricate → red at the data assertion.

Checklist

• 分支命名符合约定（fix/）
• Commit message 使用 Conventional Commits 格式
• 单一职责，改动最小化