refactor: replace 2 DB-busy retry loops with tenacity (#1132)

[axisrow]

Part of #1132 (axis 2/7 of #1130). Registry: #782 — retry batch, DB-busy group. Follows up #1198 (Counter/base64 safe subset), which explicitly deferred all retry findings for per-case owner review.

What

Replaces the two genuine hand-rolled retry loops among the 14 detector findings with tenacity.AsyncRetrying (already a direct dependency; same style as error_recovery_service.py / production_limits_service.py). The other 12 findings are poll/pacing/work-distribution loops — adjudication table below, no code changes.

1. Database._with_busy_retry (src/database/facade.py)

SQLite-busy retry on the shared write path. Behaviour preserved exactly:

• fixed delay ladder from self._busy_retry_delays_sec via wait_chain(*(wait_fixed(d) ...)) — no jitter, no exponent;
• attempts = len(delays) + 1 via stop_after_attempt;
• non-busy OperationalError re-raised immediately (retry_if_exception(_is_retryable_busy_error));
• same per-retry and final warning logs; final failure raises DatabaseBusyError chained (from) to the last busy error.

2. TelegramCommandDispatcher._update_command_safely (src/services/telegram_command_dispatcher.py)

Unbounded exponential backoff on DatabaseBusyError. Behaviour preserved exactly:

• ladder INITIAL, 2×, 4×, … capped at MAX via wait_exponential(multiplier=INITIAL, max=MAX) — bit-identical float sequence (doubling is exact in binary);
• retry_busy=False → single attempt, busy swallowed with the same log line;
• non-busy exceptions swallowed with the same log line (status updates must never kill the run loop);
• sleep=asyncio.sleep keeps the existing test seam — suite patches of mod.asyncio.sleep still observe the retry sleeps;
• CancelledError still propagates (tenacity re-raises BaseExceptions, except Exception doesn't catch them) — test_run_loop_cancelled_reraises_when_update_busy stays green.

tenacity trap discovered (for the #782 registry)

AsyncRetrying awaits the result of fn(), so fn must be an async callable. _with_busy_retry receives a sync Callable returning a coroutine (lambda: begin_immediate(...)). Without the _wrap_async adapter the coroutine was never awaited → BEGIN IMMEDIATE never ran → transaction() hit cannot COMMIT - no transaction is active, leaking out as PytestUnraisableExceptionWarning (which is an error under the project's filterwarnings = ["error"], so it silently broke every DB write and red the whole suite). The adapter is one line (async def _runner(): return await action()); the dispatcher path is unaffected because its _update is already async def. Worth a line in #782 so future retry→tenacity conversions don't repeat it.

Retry/lifecycle check (#958 trap)

Both loops are local SQLite retries — no network, no billing. SQLITE_BUSY means the statement did not execute, so a retry cannot double-apply a write. tenacity adds no hidden extra retry layer here (no SDK involved); attempt counts and delay sequences are pinned by tests.

TDD / behavioural equivalence

7 characterisation tests were added first and verified green on the old implementation, then the swap landed with the tests untouched:

• test_database.py: exact sleep ladder on exhaustion; result returned after transient busy (slept exactly delays[0]); non-busy OperationalError raise-through with zero sleeps; DatabaseBusyError.__cause__ is the last busy error.
• test_telegram_command_dispatcher.py: exponential ladder doubles and caps (computed from module constants); retry_busy=False single-shot without sleep; generic error swallowed without retry.

Detector / baseline

• Findings for both converted functions are gone; --fail-on-new green.
• Baseline refreshed (13 → 12 entries): −2 fixed; pure path renames folded in (agent/manager.py → agent/backends/_stream.py, telegram/collector.py → telegram/collector_mixins/* — same findings, files moved by earlier refactors); +1 genuinely new finding recorded (web/scheduler/context.py:236, see table).

Adjudication of the remaining 12 findings (proposal — owner decides, #782)

Finding	Class	Proposal
agent/backends/_stream.py:463 _await_with_countdown	deadline-extension ticker for streaming UX — no operation is ever re-attempted	keep (false positive)
services/task_handlers/stats.py handle_stats_all	worker-pool batch loop with flood-defer; retries delegated	keep (FP, same class as pool_dialogs #1114)
telegram_command_dispatcher.py _run_loop	infinite service poll loop (claim → dispatch); sleeps are idle pacing	keep
unified_dispatcher.py _run_loop	same	keep
collector_mixins/collection.py collect_all_channels	iterates different channels with inter-channel stagger; errors break/continue, never re-attempt	keep (FP)
collector_mixins/stats.py collect_all_stats	worker-pool with defer	keep (FP)
pool_dialogs.py warm_all_dialogs, leave_channels	already adjudicated FP — #1114, AST guards in #1149/#1176	keep (guarded)
pool_dialogs.py delete_dialogs	same pattern as leave_channels: retry delegated to run_with_flood_wait_retry, loop iterates different dialogs	keep (FP)
web/bootstrap.py _retry_telegram_pool_until_connected	condition-driven reconnect loop with cooperative-shutdown checks between waits; tenacity is exception-driven — conversion would bury the shutdown logic	keep
web/search/handlers.py _telegram_search_via_worker	poll-until-deadline for a DB command status, not a failure retry	keep (FP)
web/scheduler/context.py:236 _dedupe_recent_unavailability_events (NEW, counter)	group-by with two aggregates (count + max timestamp); collections.Counter covers only the count and would split a single-pass aggregation into two structures	propose keep

Registry row (#782 journal):

Дата	Находка (файл)	Стандартная замена	Решение	PR/issue
2026-07-10	database/facade.py _with_busy_retry; telegram_command_dispatcher.py _update_command_safely (DB-busy retry)	tenacity.AsyncRetrying	заменено (этот PR)	this PR
2026-07-10	остальные 12 retry/counter-находок	—	предложение «оставить» (таблица выше), финальное слово за владельцем	this PR

Not merging — owner reviews (retry→library is in the dual-review category per the task brief).

🤖 Generated with Claude Code