fix(dialog_cache): read cached_at as UTC-aware (naive/aware TypeError #1291)

[axisrow]

What

Fixes the latent naive/aware TypeError in the dialog-cache age computation found by audit (a "beyond scope" finding of #1234 review).

Closes #1291.

Root cause

dialog_cache.cached_at has a schema DEFAULT of datetime('now') (src/database/schema.py:316), which yields a naive string without a UTC offset. Rows inserted without an explicit cached_at therefore reached DialogCacheRepository.get_cached_at as naive datetimes, and _get_db_cached_dialogs (src/telegram/pool_dialogs.py:402) crashed on:

age_sec = (datetime.now(timezone.utc) - cached_at).total_seconds()
# TypeError: can't subtract offset-naive and offset-aware datetimes

→ the dialog warm/refresh path crashes whenever the schema DEFAULT fills cached_at. Latent only because the normal write path (replace_dialogs) emits aware isoformat() strings.

Fix

Normalize at the row→model boundary: get_cached_at now reads through parse_utc_datetime, which treats naive values as UTC. Minimal, single-method change; all consumers (pool_dialogs, web/CLI/agent dialogs handlers) benefit.

TDD cycle (per request)

• Red (proven): test_get_db_cached_dialogs_survives_naive_cached_at — real DB row relying on the schema DEFAULT raised the exact TypeError on pre-fix code in pool_dialogs.py:402.
• Green: after the fix, the e2e path and repo-level tests pass.

Tests

• tests/repositories/test_dialog_cache_repository.py (new) — schema-DEFAULT row reads back UTC-aware; normal write path preserves time without drift.
• tests/test_main_client_pool_collector_paths.py — _get_db_cached_dialogs e2e against a real DB no longer raises.

pytest tests/test_dialog_cache_ttl.py tests/test_main_client_pool_collector_paths.py \
       tests/test_telegram_client_pool_collector.py tests/repositories/test_dialog_cache_repository.py
# 282 passed

ruff check clean on changed files.

Risk / follow-up

Low. This is the recurring "naive/aware datetime at the row→model boundary" bug class (cf. #1234/pool_dialogs); normalizing on read is the project's established direction from the issue.

🤖 Generated with Claude Code

[axisrow]

📋 Review summary — all cycles

Reviewed locally in cycle-review local mode: built-in /review + Codex companion (no GitHub bots pinged).

Cycle	Reviewer	Finding	Verdict	Resolution
1	claude /review	(no findings)	—	n/a
1	codex	(no findings — verdict: approve)	—	n/a

Totals: 0 FIX, 0 SKIP, 0 UNVERIFIED. Clean first round.

Notes:

• Codex ran adversarial-review against origin/main (merge-base 0e3a0a87); verdict approve, 0 findings. Codex noted it could not execute the focused tests in its read-only env — that path is covered by the repo's own regression tests (proven red on pre-fix code).
• Round bound to head 225c88cf / base main @ 0e3a0a87.