[argus] patch #40: extract Telegram send path to _telegram_sender.py

[Nichol4s]

What

FORK-REVIEW lever 1 / stability-plan item 8. Moves the argus Telegram send path (stage-emoji indicator, HTML chunked sends with retry + plain-text fallback) out of telegram.py into argus-owned backend/app/channels/_telegram_sender.py (zero merge tax, like _telegram_format.py / _coalesce.py / _artifact_presenter.py). telegram.py keeps three thin bound-method shims (send, _send_running_reply, _send_running_reply_safe) plus its other carried patches (bytedance#17 welcome/lock, bytedance#28 webhook).

Upstream v2.0.0's superseded edit-in-place stream helpers (_send_stream_update .. _split_message, constants, _stream_messages init) are restored byte-identical as fenced, unreachable code so upstream churn in those regions merges clean instead of modify/delete-conflicting on every sync. The restored regions show zero diff vs v2.0.0.

Design constraints (why the shape is what it is)

• ALL mutable state stays on the channel instance (init_state(channel, config) from __init__) — test_telegram_send.py reads/writes ch._working_msg & friends directly.
• _send_running_reply_safe dispatches through the bound self._send_running_reply, never the module function — test_channels.py::TestTelegramProcessingOrder and test_telegram_channel_connections.py override it per instance.
• send_file is byte-identical upstream and is untouched; manager.py is untouched (its [argus] patch #46: tool_search.exclude — deferral opt-out for hot MCP tools #10 hand-off seam is already minimal and tested).

Zero behavior change — proof

• Existing suites run byte-unmodified. Full backend suite: baseline at d8ef0d1 = 5007 passed / 2 failed / 17 skipped; after this PR = 5016 passed / 2 failed / 17 skipped (+9 = the new seam tests). The 2 failures are the same pre-existing env-dependent test_artifacts_router.py xhtml cases in both runs.
• New test_telegram_sender_seam.py (9 tests) locks the shim delegation and the upstream contracts the sender consumes (OutboundMessage.progress_stage/is_final/chat_id, _application resolved at call time, Channel._on_outbound → send), so a future bad merge that resurrects upstream's send body fails loudly.
• make test-blocking-io green (21 passed); diff-scoped scan_changed_blocking_io.py reports no new candidates.

Carry impact (measured vs v2.0.0, insertions+deletions on files existing at v2.0.0)

Scope	before	after
telegram.py	574	251
app/channels/ app code	1099	776
app code excl. tests/docs	2246	1923

PATCHES.md updated in the same commit (hard rule): new bytedance#40 section, #9-chain files/delete-when reword, carry-ledger row + methodology note, and a fix to the stale bytedance#39 status note (it still said "NOT yet on argus tip" although f37b829 merged it — the d8ef0d1 ledger rebuild was authored against 2df36c9).

CI notes

• patches-ledger, backend-unit-tests, backend-blocking-io-tests should be green.
• lint-backend and frontend-unit-tests are red on pre-existing debt, unrelated to this PR. ruff check is clean on all three changed/new files; ruff format --check is clean on the two new files, and its single finding in telegram.py is a pre-existing patch-refactor: refactor the prompt management mechanism bytedance/deer-flow#17 hunk also present at the current tip (left alone to keep this PR behavior/format-pure).
• Cosmetic: log lines emitted by the moved code now carry logger name app.channels._telegram_sender instead of app.channels.telegram. No test or tooling depends on it.

After merge

Pin bump in VERSIONS.md + make deerflow-canary (live Telegram round-trip: stage emoji animates, HTML reply, indicator cleared on final) then make deerflow-promote.

🤖 Generated with Claude Code