refactor(telegram): shard adapter.py wave 1 — polling, inbound, delivery, rich mixins (−6,014 lines)

[andrexibiza]

Summary

First slice wave of the Telegram adapter god-file decomposition
(plugins/platforms/telegram/adapter.py, 10,147 lines at base be55e99cbc).
Pure extraction — zero behavior change — of the four most entangled clusters
into focused mixin modules, following the landed TelegramAuthorizationMixin
(#75742) / SessionTelegramTopicsMixin (#75747) pattern:

Module	Cluster	adapter.py net delta
telegram_polling.py (1,311 ln)	updates transport / polling / webhook	−1,241
telegram_inbound.py (1,352 ln)	inbound ingest, batching, media cache, event building	−1,275
telegram_messaging.py (1,043 ln)	outbound text delivery (send/edit/delete/draft)	−1,000
telegram_rich.py (635 ln)	rich message delivery + capability latches	−572

adapter.py 10,147 → ~4,133 lines (−6,014). Each module is imported back
into adapter.py as a module attribute (existing callers/tests stay green) and
the adapter class inherits all four mixins.

Why this wave

Operator policy: all god-files are to be sharded — model first, agree on
the graph, then slice; one PR per cluster, suite green. The decomposition plan
(three-wave blind analysis, 5 domains × 2 witnesses + validation) is the agreed
graph. This wave targeted the gnarliest clusters first (largest, most
entangled, highest regression risk) so the pattern is proven where it is
hardest. Subsequent waves take the remaining clusters (formatting, reactions,
media send/ingest, drafts, controls, notifications) with the same gates.

Verification

• Per-slice AST-fidelity (blind verifier, C:/tmp/wt-tg-verify/fidelity-report.md):
  inbound 26/26, delivery 12/12 (shim-adjusted), rich 19/19, polling verified
  in-slice with the same gates — every moved function ast.dump-identical to
  the monolith (modulo the documented function-local lazy-import shims).
• Orphan scans: zero undefined Load names left in adapter.py per slice.
• Monkeypatch-surface fix (one commit, 6012fd5e5d): the inbound mixin
  called cache_image_from_bytes via its own module global; the gateway suite
  monkeypatches plugins.platforms.telegram.adapter.cache_image_from_bytes
  (media-group batching), so the patch missed and photo-burst buffering broke
  (test_non_album_photo_burst_is_buffered_and_combined). Function-local lazy
  imports route resolution through the adapter namespace — the same shim
  pattern the delivery/rich slices use. Caught by the full-suite run, fixed,
  and regression-locked by that test.
• Suite: 59 Telegram test files on the integrated branch — 601 passed,
  0 failed (isolated basetemp; Windows host). ruff + git diff --check clean
  per slice; DCO sign-off on every commit.
• Net-shrink verified per slice from git diff numstat.

Scope honesty

Wave 1 of the adapter decomposition — the four hardest clusters. Remaining
adapter.py surface (formatting, reactions, media send/ingest, drafts, controls,
notifications, callback/picker machinery, gateway bindings) ships in follow-up
waves, each its own PR with the same evidence shape.

Links

• Decomposition plan: C:/tmp/tg-campaign/decomp-plan.md
• Blind fidelity report: C:/tmp/wt-tg-verify/fidelity-report.md
• Pattern references: refactor(telegram): extract authorization cluster into TelegramAuthorizationMixin #75742, refactor(state): extract Telegram topic cluster into SessionTelegramTopicsMixin #75747
• Prior art: lifecycle-guard god-file shard (cron), Extract Gateway Platform Routing from gateway/run.py #54962 god-file campaign

[andrexibiza]

Superseded by #79010 (adapter god-file decomposition, wave 2 consolidation). #79010's diff against main includes BOTH wave-1 (polling/inbound/delivery/rich) and wave-2 (dm-topics/lifecycle/interactive/media/config-mention/reactions) mixins — the complete 10,147 → 1,390 line decomposition in one PR. Closing this standalone to keep ONE PR per class (CCC doctrine). No work lost: every wave-1 commit is in #79010's history.

[andrexibiza]

Superseded by consolidated #79010. See epic #78791.