…ests)

Follow-up to #5228. After the first-party adapters moved into
librefang-sdk, `librefang.sidecar.adapters.telegram` shipped as a
hand-rolled stdin/stdout script that `import requests` — a dep never
declared in sdk/python/pyproject.toml (dependencies = [], no extra),
so `pip install librefang-sdk` then `python -m
librefang.sidecar.adapters.telegram` died with
ModuleNotFoundError: requests.

Rewritten as a `SidecarAdapter` subclass on the librefang.sidecar SDK,
identical in shape to the ntfy adapter, using `urllib`. The whole
`librefang.sidecar.adapters` package is now zero-dependency
(ntfy / webhook were already stdlib-only).

Transport behaviour preserved: long-poll getUpdates (server 30s /
client 35s / offset), ALLOWED_USERS numeric whitelist, sender =
first_name / username / "unknown", channel_id = chat id,
platform = "telegram", outbound sendMessage parse_mode Markdown.
Intentional SDK-consistent alignments vs the old raw script (matching
ntfy): missing TELEGRAM_BOT_TOKEN exits (supervisor reports it) rather
than emitting an error event; transport/API/send failures log + drive
the SDK's supervised backoff instead of hand-emitted protocol error
events; ready handshake is the SDK's idempotent ready/ready_ack loop.
Operator config is unchanged.

No pyproject change — the fix removes the need for `requests` entirely
rather than adding a `[telegram]` extra (keeps the adapters package
uniformly zero-dep). No Rust touched; no golden change (telegram is
not in the kernel-config schema).

Verification (clean venv, NO requests installed):
- `python3 -c "import librefang.sidecar.adapters.telegram"` — OK
- `cd sdk/python && pytest` — 32 passed: new tests/test_telegram_adapter.py
  (stdlib-only assertion; `_update_to_event` text/sender/whitelist/
  non-text-skip; `_api_get` ok / HTTPError-body / non-JSON /
  timeout-reraise; `_poll_once` emit+offset+API-error; `on_send`
  text/unsupported/skip), plus existing ntfy + SDK suites green
  (no regression).

…ment 1/N)

Promote the stdlib-only telegram adapter from a plain text relay onto
SidecarAdapter with capability negotiation, matching the in-process
crates/librefang-channels/src/telegram.rs feature surface so the kernel
routes rich commands here instead of degrading to text.

DONE this increment:
* capabilities = [typing, reaction, interactive, thread, streaming]
* reaction emoji map identical to the Rust adapter (map_reaction_emoji),
  incl. optional clear-on-done
* inline keyboards; forum-topic message_thread_id; throttled
  editMessageText streaming with UTF-16-aware 4096 chunking
  (_chunks16, the stdlib equivalent of split_to_utf16_chunks)
* inbound text long-poll, ALLOWED_USERS whitelist (unchanged)

NOT YET (gates removal of the in-process adapter, tracked):
* faithful Markdown -> Telegram-HTML formatter subsystem (this
  increment sends parse_mode=Markdown like the pre-migration adapter,
  not the Rust formatter::TelegramHtml pipeline)
* full inbound parsing of non-text updates into ChannelContent variants

Stdlib-only preserved (no requests). 37 SDK tests pass.

…inbound/outbound

Closes every NOT-YET item from increment 1, so the in-process
crates/librefang-channels/src/telegram.rs is now functionally
superseded by the sidecar (its removal is the separate kernel-cascade
change, deliberately not bundled here).

Faithful, function-by-function ports of the audited Rust (not
re-derived):

* Markdown -> Telegram-HTML formatter: byte-exact port of
  formatter::markdown_to_telegram_html (+ render_inline / fence /
  heading / list helpers), pinned by tests to the Rust
  formatter::tests oracle (bold/italic/code/link/heading/list/
  fenced-code/blockquote).
* sanitize_telegram_html security pass: tag + href-scheme allowlist,
  attribute-injection escaping, unclosed-tag balancing, rebuild_safe_tag
  / parse_attrs / is_safe_href — including the Rust edge where a
  dropped unsafe-<a> leaves its </a> escaped.
* message_truncator port: utf16_len / truncate_to_utf16_limit /
  split_to_utf16_chunks with the HTML-entity boundary guard and
  zero-progress guards.
* Outbound text now formats + sanitizes + chunks and sends
  parse_mode=HTML with the same plain-text retry on Telegram's
  "can't parse entities" 400 (was a naive parse_mode=Markdown).
* Full inbound: text/bot-command, photo, document, audio, voice,
  animation, video, video_note, location, sticker; from/sender_chat
  sender extraction; callback_query -> ButtonCallback (+ fire-and-forget
  answerCallbackQuery); edited_message; reply-to context; getFile URL
  resolution with text fallback; ALLOWED_USERS by id and username.
* Full outbound ChannelContent dispatch for every variant
  (Image/File/Voice/Video/Audio/Animation/Sticker/Location/MediaGroup/
  Poll/Interactive/EditInteractive/DeleteMessage), private-URL ->
  multipart upload, OGG/Opus voice routing, 429 retry_after.

Stdlib-only preserved. Full Python SDK suite 43 pass (telegram
parity + cross-impl conformance + ntfy/webhook/runtime, no regression).