fix: bound the inline state map, test the 300s sweep, add a bot-wide send budget

Three gaps the last audit list named, all in the "resource growth, background
timers and limits nobody watches" class.

**Idle inline-query entries are pruned.** `DebounceStates` had no eviction at
all: one entry per user who ever used inline mode, forever, while the rate
limiter's buckets and the chat store both prune in the 300s sweep. Entries
now carry a `last_seen` stamp and `prune_idle_states()` drops the ones idle
past 300s — the window Telegram caches an inline answer for
(`cache_time(300)`), after which a repeat reaches the bot again and has to be
answered fresh, so the entry would only suppress a fetch the user is waiting
for. The boundary is tested through `prune_idle_at(now, idle_for)` so it does
not depend on ageing a monotonic clock.

**The 300s sweep is a function, and tested.** It was an inline `tokio::spawn`
block: the expiry edit (the only part that talks to Telegram) had no test at
all. It is now `periodic_sweep(sender, chat_store, link_cache, task_queue,
config, stop)`, which also prunes the inline entries, driven in a test with
`start_paused` — the loop's own timer fires the tick, exactly one expired
prompt is rewritten in place, a live one keeps its record and buttons. The
interval is pinned as a constant because no assertion on the edits can see it
(a shorter one produces the same single edit; the paused clock can jump past
the boundary while a tick's DB work is in flight). To make the edit reachable
at all, `edit_message_text` joined the `MediaSender` trait (Bot impl + mock
recording), which is also what keeps `main.rs`'s remaining `Bot` calls
unambiguous. `main.rs` leaves the "untested modules" list except for
startup/shutdown and the dispatcher tree.

**The bot-wide send budget exists.** Telegram throttles a bot in total
(~30 msg/s) as well as per chat; only the per-chat bucket existed, so a batch
forward fanned out over many chats was unguarded and earned 429s the queue
then retried. `acquire_global` charges the same spend against a single shared
bucket at the three paced sites (`send_media_group`, `send_animation`,
`copy_messages`). The unpaced ones (`send_message`, the edits, the toasts) stay
unpaced on purpose: they are one call per action, far below the ceiling, and
pacing a user-visible reply would delay it. Not covered: that the send paths
call it (they need a real `Bot`), which is the same structural gap as the
dispatcher tree.

Also: the startup token-exchange decision is now `startup_validation(result)`
instead of living inside the `Site::validate` future, so "a 5xx while the
container comes up must not disable pixiv" is asserted as a decision — the
message the admin gets plus `enabled()` unchanged. The rejected-credential
half is deliberately not exercised: it calls `disable()`, a process-wide flag
with no reset, and a test touching it would order-couple every other pixiv
test.

Verified: `cargo fmt`, `cargo clippy --workspace --all-targets --locked -- -D
warnings`, `cargo test --workspace --locked` (184 passed, 14 ignored) — plus
mutations, each confirmed to fail the relevant test: the sweep not being
driven on its timer, the interval shortened to 60s, and (earlier) the queue
sweep's missing wake-up. Dropped an empty leftover `crates/x-media/tests/`
directory while there (never tracked by git).

AGENTS.md

@@ -38,7 +38,7 @@ The `x-media` library: `site::fetch(url)` dispatches through the `SITES` registr
 |---|---|
 | `crates/x-media/src/` | Fetch library. `site/mod.rs` = dispatcher + `Fetched`/`FetchError`/`download_media`/`media_size`; `media.rs` = `Media` enum; `examples/fetch.rs` = end-to-end usage sample |
 | `crates/x-media/src/site/<twitter\|pixiv\|bsky\|misskey\|bilibili>/` | One directory per site: `mod.rs` (re-exports), `interface.rs` (PATTERN, `enabled()`, `fetch_from_url()`, `cache_key`/`is_retryable`/`media_headers`, unit struct `<Name>Site` implementing `site::Site`, `From<SiteStruct> for Fetched`), `model.rs` (serde DTOs). Pixiv adds `api.rs` (auth + transport); twitter adds `auth.rs` (logged-in GraphQL `TweetDetail` fallback for NSFW tweets, gated on `TWITTER_AUTH_TOKEN`; without the token a withheld tweet stays `FetchError::Sensitive` and the bot reports it as age-restricted instead of "no media"). Misskey targets misskey.io only (`POST /api/notes/show`, 400+`NO_SUCH_NOTE` → NotFound). Bilibili fetches dynamics (images/animated images only — an attached video degrades to its cover, and its title stands in for the post text, which AV dynamics do not have) from `/x/polymer/web-dynamic/v1/detail` sent with `features=itemOpusStyle` (without that flag the legacy serialization drops an image/text post's body and headline entirely — `desc` comes back `null`; the adapter still parses the legacy `major.draw`/`desc`/`archive` shapes as a fallback). No WBI signature is involved; device cookies `buvid3`/`buvid4` are fetched automatically from `/x/frontend/finger/spi` because bilibili's `-352` risk control starts rejecting plain requests, `BILIBILI_COOKIE` is the escalation when an IP stays blocked; `b23.tv` short links are deliberately unmatched. Twitter's `from_syndication_json` HTML-decodes the API text — syndication and GraphQL `full_text` both arrive pre-escaped (`&gt;` `&lt;` `&amp;` `&#39;`) — so the stored text is raw and the caption escap…
-| `crates/xmedia-bot/src/main.rs` | Entry point: env/log init, command registration (`register_commands` — `setMyCommands` plus the profile description texts), shared `send::BOT` force-init, startup sweep of this project's leftover temp files (`x_media::TEMP_FILE_PREFIX` + an age gate, since a killed process runs no destructors), queue worker start, site login validation (`site::validate_all`), 300 s edit-expiry sweep (expired prompts are rewritten in place to `EDIT_PROMPT_EXPIRED_TEXT` with an empty keyboard — an edit, never a new message, so a background timer cannot wake a chat), dptree handler tree, webhook vs polling dispatch |
+| `crates/xmedia-bot/src/main.rs` | Entry point: env/log init, command registration (`register_commands` — `setMyCommands` plus the profile description texts), shared `send::BOT` force-init, startup sweep of this project's leftover temp files (`x_media::TEMP_FILE_PREFIX` + an age gate, since a killed process runs no destructors), queue worker start, site login validation (`site::validate_all`), `periodic_sweep` (`SWEEP_INTERVAL` 300 s): expired prompts are rewritten in place to `EDIT_PROMPT_EXPIRED_TEXT` with an empty keyboard — an edit, never a new message, so a background timer cannot wake a chat — plus the link-cache prune, the idle rate-limit buckets and the idle inline-query entries, and the queue backlog line (only when non-empty). Takes its collaborators rather than the statics so its loop is testable with a paused clock, dptree handler tree, webhook vs polling dispatch |
 | `crates/xmedia-bot/src/config.rs` | Manual env parsing into `Config` |
 | `crates/xmedia-bot/src/db.rs` | `DbPool`: one shared SQLite connection pool (`POOL_SIZE = 4`, WAL, busy_timeout) for all three tables over `$DATA_DIR/task_queue.db` (default `data/`) — the three stores share it; `open_store` creates file + schema and then applies the `PRAGMA user_version` migration chain (`MIGRATIONS` + `migrate` — append-only; `schema_init` is the version-0 baseline and must not gain columns an existing database would never receive), `with_conn` runs all rusqlite I/O in `spawn_blocking` |
 | `crates/xmedia-bot/src/handlers/` | Handler modules: `mod.rs` (message entry point, `reply`, `log_key`, the group-only `GROUP_LINK_HINT` for a supported link posted outside a private chat), `commands.rs` (teloxide `BotCommands` enum + command executor, incl. `/test <url>` (send-only) / `/debug <url>` (parse-only) and the admin-only `/bot_dict` state dump; `/set_format` rejects unknown `{…}` placeholders and resets with `-`), `urls.rs` (URL extraction + bounded job channel (256) drained by `URL_WORKERS = 8` workers (`start_url_workers`) — backpressure instead of unbounded spawns; teloxide's per-chat workers are sequential — batch-forwards need concurrency), `inline.rs`/`callback.rs` (inline queries / edit-before-forward buttons, incl. `skip`; a forward that fails retryably is both queued *and* settles the prompt — the queued row carries the message ids itself, and a prompt left live let a second Confirm copy the same messages twice and let Skip answer "nothing was forwarded" while the row still delivered), `statics.rs` (global statics) |
@@ -47,8 +47,8 @@ The `x-media` library: `site::fetch(url)` dispatches through the `SITES` registr
 | `crates/xmedia-bot/src/queue.rs` | `PersistentTaskQueue`: SQLite-backed queue (`tasks` table), `QUEUE_WORKERS = 4` concurrent workers (lease via `BEGIN IMMEDIATE` + `locked_until` TTL), retry→dead-letter, a `lease_token` fence: `lease_next` stamps a random token and every write-back (heartbeat, `delete`, `reschedule`, `mark_done`) is guarded by it, so a lease that expired and was re-leased cannot be written by its former holder — a lost lease stops the attempt instead; a finished row's `DELETE`/reschedule retried and a failed delete falling back to a `done` tombstone (the lease query and the sweep only look at `pending`/`in_progress`, so a task that already ran cannot be resurrected and re-run), `notify_one` worker wakeup plus a separate `Notify` for the 30 s lease-expiry sweep (a shared one let the sweep steal the workers' wakeup permit; the sweep does notify the workers after it actually recovered a row, since a recovered task is due immediately while every worker may be parked on `notify` with no pending row to sleep on), `busy_timeout` on all connections |
 | `crates/xmedia-bot/src/ctx.rs` | `AppContext`: the injected collaborators (`sender` + `ChatStore`/`PersistentTaskQueue`/`LinkCache`/`Config`), `from_statics` for production and the `CONTEXT` static the worker closures hold. `test_support::TestStores` backs handler tests with a tempdir store set, and the module also carries the fixtures those tests share — the canonical cached post (`cached_photo`), the edit-before-forward prompt (`seed_prompt` with its `PROMPT_ID`/`FORWARDED_ID`) and a scripted API error (`api_error`) — so no two test modules keep their own copies |
 | `crates/xmedia-bot/src/send/` | `send/mod.rs`: `Task`/`MediaItemPayload` payloads, `SendError`/`Classification`, `send_media_sequence`/`send_animation`/`forward_messages`; `send/input_media.rs`: payload → `InputFile`/`InputMedia` + `build_media_group` (caption on the first item only); `send/upload.rs`: the download-and-reupload fallback (`prepare_upload_item`/`send_batch_via_upload`, photo downscale handoff); `send/post_send.rs`: link-cache write, `KEEP_ALIVE` registry, `settle_task`, `post_send_actions`, `handle_task`/`dead_letter_notify` |
-| `crates/xmedia-bot/src/media_sender.rs` | `MediaSender` trait: the user-flow surface (`send_media_group`/`send_animation`/`copy_messages`/`send_message`/`answer_callback_query`/`edit_message_caption`/`delete_message`/`send_chat_action`) implemented by teloxide `Bot` (per-chat rate-limited) and by a recording `MockSender` in tests. Admin/setup APIs (`get_chat`, `set_my_commands`, …) stay on the concrete `Bot` |
-| `crates/xmedia-bot/src/rate_limit.rs` | Per-chat token bucket (`CAPACITY = 20`, ~20 msg/min refill) paced before sends reach the API so batch forwards don't trip flood control |
+| `crates/xmedia-bot/src/media_sender.rs` | `MediaSender` trait: the user-flow surface (`send_media_group`/`send_animation`/`copy_messages`/`send_message`/`answer_callback_query`/`edit_message_text`/`edit_message_caption`/`delete_message`/`send_chat_action`) implemented by teloxide `Bot` (per-chat rate-limited) and by a recording `MockSender` in tests. Admin/setup APIs (`get_chat`, `set_my_commands`, …) stay on the concrete `Bot` |
+| `crates/xmedia-bot/src/rate_limit.rs` | Two token buckets paced before sends reach the API so batch forwards don't trip flood control: one per chat (`CAPACITY = 20`, ~20 msg/min refill) and one bot-wide (`acquire_global`, 30/s — Telegram's per-bot ceiling, invisible to any per-chat bucket and only binding when a batch fans out over many chats). `prune_idle` drops the per-chat buckets that refilled while unheld |
 
 ## Development Commands
 
@@ -110,5 +110,5 @@ Docker: `docker build -t tgxmb .` then `docker run --rm -d --name tgxmb --env-fi
 - Live-network tests exist in `site/twitter/interface.rs` (5), `site/bsky/interface.rs` (1), `site/misskey/interface.rs` (1), `site/bilibili/interface.rs` (5), `site/pixiv/api.rs` (1); `photo.rs` adds one `#[ignore = "heavy: …"]` test. `site/mod.rs` also has a **token-gated but not `#[ignore]`d** pixiv download test (`download_media_pixiv_original_with_referer`): it hits `i.pximg.net` whenever `PIXIV_REFRESH_TOKEN` is set, so a local `cargo test --workspace` is not fully offline and can flake on a pixiv CDN body timeout. `disabled_site_is_reported_not_ignored` (same file) is gated the other way round: it asserts `fetch` answers `FetchError::Disabled { site: "pixiv" }` for a pixiv link and early-returns when `PIXIV_REFRESH_TOKEN` **is** set (the site is then enabled). Test gating convention (enforced by `.github/workflows/ci.yml`): pure unit tests always run; live-network tests carry `#[ignore = "live network: ..."]` (run via `cargo test --workspace -- --ignored live`); token-gated pixiv tests early-return when `PIXIV_REFRESH_TOKEN` is absent **or empty** (an unset GitHub secret arrives as `""` — `is_err()` alone would run them tokenless and fail), and the bilibili live tests early-return when the API answers risk control (`-352`, which bilibili applies per IP by request volume). Run the full offline suite with `cargo test --workspace`.
 - Fixtures are inline `serde_json::json!` builder fns (`fixture()`, `thread_json()`, `illust_json()`), not files. The shared `CLIENT` sets `pool_max_idle_per_host(0)` under `#[cfg(test)]` to avoid cross-runtime `DispatchGone`.
 - **CI** — `.github/workflows/ci.yml` (actions pinned to commit SHAs, `--locked` on every cargo invocation, `concurrency` cancels superseded runs, `RUST_BACKTRACE=1`) runs `cargo fmt --check` + `cargo clippy --workspace --all-targets --locked -- -D warnings` + `cargo test --workspace --locked` + a release-profile `cargo build --release --locked` + an `actions-rust-lang/audit` dependency-vulnerability gate (offline, no secrets, on every push/PR) and a `live` job (schedule/manual/tag only, `-p x-media` since every network/secret-gated test lives there, `continue-on-error`) for the `#[ignore]`d live + token tests. `.github/workflows/docker.yml` builds and pushes the image on master/tag and runs a **build-only check on pull requests touching the build inputs** (`Dockerfile`, entrypoint, manifests, `.dockerignore`); a release tag must match both crate versions or the build stops, and `FFMPEG_URL`/`FFMPEG_SHA256` are taken from repository variables when set (a release can pin an exact ffmpeg build). `.github/dependabot.yml` keeps crates, the pinned actions and the Docker base images current.
-- Untested and hard to test without a mock seam: `main.rs`, `config.rs`, `db.rs`, `handlers/statics.rs`, `media_sender.rs` (holds the `MockSender` itself); in `x-media`: `media.rs`, `lib.rs`, all `model.rs`. The `commands.rs` *executor* needs a real `Bot` (only its pure report builder is tested). Everything else — `handlers/{mod,callback,inline,urls}.rs`, `send/*`, `ctx.rs`, `state.rs`, `queue.rs`, `link_cache.rs`, `rate_limit.rs` — is driven through `TestStores`/`ctx::test_support` and the scripted `MockSender`.
+- Untested and hard to test without a mock seam: `config.rs`, `db.rs`, `handlers/statics.rs`, `media_sender.rs` (holds the `MockSender` itself, so only the trait's mock side is exercised); `main.rs` is covered where it was split out (`periodic_sweep`, `sweep_temp_dir`) but not for startup/shutdown or the dispatcher tree; in `x-media`: `media.rs`, `lib.rs`, all `model.rs`. The `commands.rs` *executor* needs a real `Bot` (only its pure report builder is tested). Everything else — `handlers/{mod,callback,inline,urls}.rs`, `send/*`, `ctx.rs`, `state.rs`, `queue.rs`, `link_cache.rs`, `rate_limit.rs` — is driven through `TestStores`/`ctx::test_support` and the scripted `MockSender`.
 - No coverage tracking.