Real-Telegram test suite: 11 -> 32 tests, every command and button covered

[Szer]

Expands the real-Telegram (MTProto) suite from 11 to 32 tests, covering every user-facing
command and every inline button the bot builds — happy paths only, since the hermetic suite
(192 tests) owns the edge cases.

Real-run evidence: https://github.com/Szer/bots/actions/runs/30194699791 — Passed: 32, 9m56s.
Hermetic suite unchanged at 192/192. No production code changed.

Coverage added

Before this PR, the real suite drove 4 inline buttons. It now drives every callback the bot
emits, and every command including aliases.

New file	Tests	Covers
AddWizardRealTests	6	the whole interactive /add wizard — `addflow:ocr:yes
MyAndAddedRealTests	7	/added + void:<id>, myAdded, the bare used:<id>/return:<id> buttons, and text /used, /return, /void
ReportButtonFlowRealTests	4	/my → report → report:<id> → report:<id>:confirm, reportCancel at both steps, reportedUsed:<id>
AdminAndFeedbackRealTests	4	/feedback end-to-end, admin /debug and /undo, and all six command aliases

Two gaps worth naming: /report previously had coverage only as a text command, while the button
chain — the path announced to the community — had none. The /add wizard, the most-used and most
stateful add path in production, had none at all.

Shared plumbing

• RealTestHelpers.addCouponViaCaptionAsync — the idempotent add: clears the fixture's barcode,
  sends the photo, returns the new coupon id.
• DbSeed gains deleteCouponsByBarcodeAsync, deleteCouponByIdAsync, releaseTakenCouponAsync,
  getCouponStatusAsync, getLatestOwnedCouponIdAsync, setCouponExpiryAsync,
  deletePendingAddFlowAsync, deletePendingBatchesAsync.
• TgUserClient gains PressCallbackButtonMatching and AwaitAndPressButton; a missing button now
  reports which callback datas were actually present instead of failing opaquely.
• tests/CouponHubBot.RealTests/README.md documents the skeleton, the helpers, and the idempotence
  rule below.
• CI job timeout-minutes 30 → 50; 32 tests plus the ~7-minute deploy chain no longer fit in 30.

The one rule this suite lives by

Every coupon add must be idempotent. 32 tests share one database, one Telegram account, and one
bot, with no per-test cleanup and no guaranteed class order. Since #269 made manual /add extract
barcodes, adding a fixture whose barcode already exists is rejected with «Купон с таким штрихкодом
уже есть в базе…» rather than «Добавлен купон» — so a leftover row from any other test silently
breaks an unrelated one.

"The other test deletes first, so we're fine" is not a safety argument: if the deleting test runs
first, it leaves the row behind and the non-deleting test is the one that fails. Every add site now
clears its own barcode immediately before sending. This is stated at the top of the suite README.

Three CI dispatches were needed to get here (27/32 → 28/32 → 32/32). Every failure across both red
runs came from shared mutable state, not from bot bugs and not from latency:

• non-idempotent adds in the pre-existing tests colliding with the new ones;
• MAX_TAKEN_COUPONS=6 exhausted by tests that left coupons taken forever, after which /take
  replies with a text limit message that no photo-caption await can ever match;
• ReminderRealTests inserting a coupon with a synthetic photo_file_id, which made /my fail its
  whole media-group send with wrong remote file identifier;
• bulk-cancel confirming by editing its message, while the MTProto client deliberately drops edit
  updates — so the awaited text could never arrive;
• Dapper 2.1.72 silently returning default(DateOnly) for a bare scalar DateOnly query.

Notes for future readers

• addflow:ocr:yes needs a SQL bump of the staged expiry, permanently. Fixtures must stay
  expired coupons because this repo is public, so OCR always reads a past date, and TryAddCoupon
  rejects expires_at < today before any DB write. There is a doc comment at the site; it is not a
  temporary hack to be cleaned up.
• BulkAddRealTests still overrides expires_at via direct SQL for the same reason, so it does not
  exercise the batch path's expiry validation.
• The suite remains sensitive to real-Telegram latency. TestRetry retries once, on timeout-class
  failures only — never on assertion failures.

🤖 Generated with Claude Code

https://claude.ai/code/session_01P41roCWN5Fec7m69sSaunU