fix(admin): state the refusal a deleted model now meets, not the old zero bill

[argszero]

Summary

admin.models.sub — the card subtitle on the admin models page, in both packs and the index.html static fallback — tells admins that "calls for a deleted model bill at 0". So do four code comments (admin_models' module doc and its remove doc, db::seed_models, sharing::create and its C2065 test).

That was true when the sentence was written (98a7a9b, v0.6.4). R156 (acb8815) closed that state: forward and forward_stream now price every candidate key before picking one (src/gateway.rs:278-296 / :512-523), so a status='on' key whose (provider, model) has left the catalog is dropped from the pick set and the call is refused with 503 ("provider 与 model 不在模型目录中，无法计价") before anything reaches upstream. The settlement fallback arm is a loud log::error! now, not a silent zero.

R156 touched three files (db.rs, gateway.rs, state_gate.rs) and left every carrier of the old sentence in place — including db.rs, where it rewrote the first half of that comment into the invariant and kept the stale half.

One consequence worth naming: the card subtitle is the only place an admin meets this rule. The catalog editor is exactly where the state is created (deleting or renaming a catalog row does not touch keys — admin_models::remove is a bare DELETE FROM models WHERE id = ?1), so the page that creates the condition is also the page that misdescribes it.

Related Issue

None — found during this task's own reconnaissance. No open issue covers it (gh issue list --state open is empty).

Changes

• Fix description: admin.models.sub (zh + en in ui/js/i18n.js, and the ui/index.html static fallback, kept byte-identical to the zh value) now states the refusal instead of the zero bill.
• The wording is not new: it reuses the "无法计价 / cannot be priced" family the refusal itself already carries (err.modelNotInCatalog, the ERR_MAP entry derived from the backend literals in routes::{gateway,sharing}), so the copy and the error the user actually gets speak with one voice.
• ui/js/i18n.js?v=20260927-2 → ?v=20260927-3 per the cache-bust convention in ui/README.md.
• Four comments now point at the authority (the pricing gate at the routing entry) instead of restating a rule that lives in code — including admin_models' module doc, which already claimed not to restate it.
• Configuration/data-structure changes synced with the example file: n/a — no config, no schema, no API change.

What is deliberately not changed

• CHANGELOG.md:243 (the v0.6.4 entry, "deleted models billed at 0"). It was true when written; this repository does not rewrite history entries (unlike the v0.7.28 correction, where the token was wrong at the moment it was typed).
• src/gateway.rs in full, and the R156 gate header in src/state_gate.rs. Those passages describe why the gate exists — their subject is the defect that was closed, not today's behaviour. Rewriting them would erase the reasoning that justifies the code, and the gate's R2 fixture (state_gate.rs:17111) quotes the fallback arm's log::error! literal verbatim.
• The settlement fallback arm's own message ("⇒ 本次按 0 计费"): inside that arm it is still accurate — that arm bills zero, which is exactly why it logs.

Why no new gate

This is the doc-comment-claims axis — a claim that stopped being true — where this repository's settled answer is to fix the data, not add a lexical guard (C2024 d2c7431, C2025 9db4872, C2059 81df12d; most recently the R89 45510ba and R91 194cc5f rounds). A lexical gate can pin the shape of a sentence, never its truth, and both mechanical alternatives are worse than the defect they would guard: making the copy quote the refusal sentence would mint a second carrier of one fact (the thing this repository spends its gates removing), and a hand-written list of forbidden words is a wordlist, not an invariant.

Tests

• Existing tests pass — cargo test: 413 passed / 0 failed (baseline on 11aaba5 is 413; 413 filtered out-style counts unchanged).
• cargo fmt --check — clean.
• cargo clippy --all-targets -- -D warnings — clean.
• New tests added: n/a (see "Why no new gate") — the affected i18n counters (ZH/EN_KEY_COUNT, T_LITERAL_*, STATIC_ATTR_*, ERR_MAP_ENTRY_COUNT) are untouched by this change and keep passing unchanged, which is the evidence that the edit added no key, no literal and no attribute.

Checklist

• Branch naming follows the convention (fix/…)
• Commit message uses Conventional Commits
• Single responsibility, minimal change (5 files, +19/−18, no production logic touched)

[argszero]

Self-review (this task holds allow_self_merge: true; GitHub refuses the formal approve button on one's own PR, so the record lives here as an ordinary comment).

Does the diff say what the base commit does? Yes. gateway::forward and forward_stream price every candidate key before router.pick and refuse with 503 when none can be priced (src/gateway.rs:278-296 / :512-523), and gateway::tests::unpriced_model_503_before_upstream pins exactly that: a key whose (provider, model) is absent from models but still routed by model gets SERVICE_UNAVAILABLE, the message contains 无法计价, and cooldown_len() == 0 — nothing reaches upstream. admin_models::remove is a bare DELETE FROM models WHERE id = ?1, so the state the old copy described is still creatable — which is precisely why the copy has to describe what happens next, and it now does.

Is the new wording ours or newly minted? Ours. It reuses the err.modelNotInCatalog family ("无法计价" / "cannot be priced") that the 503 itself carries and that ERR_MAP maps from routes::{gateway,sharing} literals — so the card subtitle and the error a user actually meets speak with one voice.

Did anything shift? No key, no literal, no attribute: ZH/EN_KEY_COUNT 796, T_LITERAL_DISTINCT 435, STATIC_ATTR_DISTINCT 309, ERR_MAP_ENTRY_COUNT 52 all pass unchanged.

Deliberately absent: a new gate. This is the doc-comment-claims axis and the repository's settled answer is to fix the data (C2024 / C2025 / C2059, plus R89 45510ba and R91 194cc5f in this same series); the two mechanical alternatives both make it worse (see the PR body).

Verified on the branch: cargo test 413 passed / 0 failed (baseline 413), cargo fmt --check clean, cargo clippy --all-targets -- -D warnings clean. CI run 36328012923: both jobs green (test / fmt / clippy, msrv).