fix(telegram): discard buffered ingress fragments at conversation boundaries

[Enough1122]

Fixes #81370

Root cause

Telegram keeps per-session debounce slots for inbound text batching, photo bursts/albums, and media-group coalescing (implemented for #46100 / #46101 in plugins/platforms/telegram/adapter.py):

• _pending_text_batches / _pending_text_batch_tasks — _enqueue_text_event
• _pending_photo_batches / _pending_photo_batch_tasks — _flush_photo_batch
• _media_group_events / _media_group_tasks — _queue_media_group_event

Only disconnect() (_cancel_pending_delayed_deliveries) cleared them. A conversation boundary (/stop, /new, /reset, auto-reset, expiry finalization) never triggers disconnect(), so a forward or media download that finished after the reset kept its fragments in those slots — and the next unrelated user message merged them.

The production evidence from the issue: forwarded video 411357 was still in the startup/photo slot after /stop (411367); 2+ minutes later, ordinary copied text 411376 merged it and the gateway logged Merged 1 Telegram startup attachment(s) for a turn where the user sent no media.

Fix

1. New hook BasePlatformAdapter._discard_session_ingress(session_key) in gateway/platforms/base.py (default no-op), funneled through _discard_text_debounce, which every conversation boundary already calls via cancel_session_processing and the /stop,/new,/reset bypass-dispatch path.
2. TelegramAdapter._discard_session_ingress overrides it: drops the session's text-batch, photo-burst/album, and media-group slots and cancels their flush tasks. Unrelated sessions' buffers are keyed differently and survive untouched.

The fix is session-scoped — one conversation's reset cannot affect a concurrent session's in-flight batch.

Regression tests

5 tests in tests/gateway/test_telegram_ingress_boundary_81370.py:

1. test_text_batch_for_session_is_dropped — the issue's scenario: a stale text-batch fragment is gone after the boundary.
2. test_other_sessions_batches_survive — session scoping: another session's batch is untouched.
3. test_photo_burst_and_album_slots_for_session_are_dropped — photo burst (session:photo-burst) and album (session:album:<id>) slots both cleared.
4. test_media_group_slots_for_session_are_dropped — media-group coalescing keyed by media_group_id is matched via the buffered event's source.
5. test_base_hook_funnel — _discard_text_debounce funnels into the adapter override, so every existing boundary that clears the base debounce also clears Telegram's slots.

All 5 are RED on pre-fix code (hook missing → AttributeError) and GREEN after. Existing Telegram tests (test_telegram_caption_merge.py, test_telegram_auth_check.py, test_telegram_audio_vs_voice.py, test_telegram_approval_buttons.py — 27 tests) continue to pass.

Scope

One new base hook (no-op default) + one adapter override + tests. No refactor of the debounce machinery itself; the existing boundary funnel is reused rather than adding a parallel cleanup path.

[alt-glitch]

This was generated by AI during triage.

Related to #81371, which addresses #81370 with a broader generation-scoped ingress design. This PR uses explicit session-boundary cancellation and also includes an unrelated oneshot fallback change for #81209; please consolidate the overlapping Telegram fix before merge.

[Enough1122]

Critical fix: the media-group branch of _discard_session_ingress matched event.source.session_key == session_key, but SessionSource has no session_key attribute — the session key is always derived via build_session_key(source). The lookup returned None for every real source, so the media-group cleanup branch was dead code and Telegram media-group albums were never invalidated at conversation boundaries (the exact scenario the PR title targets). The existing test masked the defect with a fabricated source fixture.

This follow-up:

• Recomputes the key via build_session_key(event.source, ...) (with try/except) so the cleanup matches real SessionSource objects.
• Rewrites the regression test to use a real SessionSource (Platform.TELEGRAM) so this branch can never silently regress.

5/5 tests pass locally.

[Enough1122]

Closing in favor of #81371 (Qwinty), which was opened ~6h earlier (2026-08-07T23:20Z vs 08-08T05:24Z) and fixes the same issue #81370 (stale Telegram ingress fragments leaking across conversation boundaries) in the same files (gateway/platforms/base.py + plugins/platforms/telegram/adapter.py). Our 5 regression tests in tests/gateway/test_telegram_ingress_boundary_81370.py remain on branch fix/81370-telegram-attachment-buffer-leak if maintainers want to salvage any of them.