perf(http): route Telegram bot service and TTS client through the shared pool (#12979)

[mrveiss]

Thinking Path

Batch 5 of #12979. Per the consolidated batch recipe on the issue: classify every site by reading it (do not grep-and-sweep), because the *_json() vs tracked_request() split has swung 82%/0%/72%/0% across the first four batches and is a property of the module's contract, not the codebase.

Scope: the two dense single files named in the task — autobot-backend/services/telegram_bot_service.py (8 sites) and autobot-backend/services/tts_client.py (7 sites) — 15 sites total, at the ~15 target, so no third area was added.

Read every site in both files:

• Telegram (8/8 tracked_request()): every method reads error_text = await response.text() on the non-2xx branch before raising, specifically to log the response body — get_json()/post_json() would raise ClientResponseError immediately and drop that log detail, a behaviour change in exactly the place batch 3's _trigger_build() finding warned about (raise_for_status-shaped code that isn't a *_json() candidate once you look at what it actually reads). download_file() reads raw bytes(). verify_token() inspects status == 200 and returns a bool sentinel, never raising.
• TTS client (7/7 tracked_request()): is_available()/list_voices()/delete_voice() swallow failure into a debug/warning log plus a sentinel (bool/[]); synthesize()/clone_voice()/create_voice() read raw bytes() or raise a custom RuntimeError with the response body already read via text(); synthesize_stream() streams length-prefixed chunks from resp.content.readexactly() inside the response block.

No site in either file uses raise_for_status(), a custom connector=, returns a session, or is long-lived — so this batch has zero RAW carve-outs, unlike batch 4's SSRF-pinned connector.

synthesize_stream() is a genuine streaming case: the tracked_request() async with block spans every yield in the generator body. This works because async with remains active across await/yield points in an async generator — the block exits (and the counter decrements) whichever way the generator ends, exhausted normally or closed early by the caller giving up on the stream.

What Changed

• autobot-backend/services/telegram_bot_service.py: 8 raw aiohttp.ClientSession() → 0, all via get_http_client().tracked_request(method, url, ...). Removed the now-unused import aiohttp.
• autobot-backend/services/tts_client.py: 7 raw aiohttp.ClientSession(timeout=...) → 0, all via tracked_request(). is_available(), list_voices(), and delete_voice() pass suppress_error_log=True since each already downgrades the failure to a debug/warning log plus a sentinel — the pooled client's default ERROR log would otherwise turn an expected condition (TTS worker not running) into noise on a poll loop.
• autobot-backend/services/tts_client_test.py: retargeted from patch("aiohttp.ClientSession", ...) to patch("services.tts_client.get_http_client", ...), since the code no longer touches aiohttp.ClientSession at the call site. The fake client's tracked_request() returns the same async-context-manager mock the old fake session's .post() returned.
• autobot-slm-backend/tests/services/tts_contract_test.py: a second, non-mock test seam, found by exercising rather than by the repo-wide patch(.*ClientSession grep. This is a static regex contract test (bug(tts): synthesis 404s — deployed worker lacks /tts/synthesize/stream because autobot-tts-worker has NO resolve path AND NO drift visibility (silent skew on every backend update); /tts/clone-voice is a dead call that exists nowhere #12886) that parses tts_client.py's source for the session.<verb>(f"{self.base_url}/path") shape to confirm every backend call is served by the worker template. The regex didn't match the new tracked_request("VERB", f"{self.base_url}/path", ...) shape and went 8 failed / 1 passed. Added a second regex for the pooled shape alongside the original (not a replacement), so a future raw-session call site is still caught by the same test.
• telegram_bot_service.py has no direct unit test exercising its HTTP methods (api/telegram_bot_test.py covers webhook wiring and command handling only), so there was no seam to retarget there.

Verification

(a) py_compile, flake8, black --check --line-length=120 — all clean on the 4 changed files.

(b) Local aiohttp server exercise — all 15 converted sites driven on success and failure branches against a real local server (29 exercised paths, 0 unexpected):

exercised paths       : 29
unexpected            : 0  []
server-side requests  : 29
still-acquired        : 0
active_requests       : 0
ALL ASSERTIONS PASSED

(c) Baseline vs after (cp-swapped origin/Dev_new_gui, never git stash) — repo-wide grep -rn 'patch(.*ClientSession' confirmed the only mock-based seam for these two modules is tts_client_test.py (fixed above); the regex-based seam in tts_contract_test.py was caught by running it, not by that grep, which is why "exercise every path" and "diff against baseline" both matter independently. Curated regression scope (both files' direct tests plus every caller's test file: api/telegram_bot_test.py, api/voice_stream_test.py, tests/test_voice_503_guard.py, integrations/protocols_conformance_test.py, tests/services/notification_service_test.py, autobot-slm-backend/tests/services/tts_contract_test.py):

baseline (origin/Dev_new_gui): 119 passed, 0 failed
after this branch:             119 passed, 0 failed

A full autobot-backend/ suite run was attempted for both states but hung indefinitely on a pre-existing, unrelated security/security_api_e2e_test.py before reaching services/ alphabetically — confirmed pre-existing since it hung identically with the baseline (origin) files in place. The curated scope above stands in for it; it covers every known caller of the two converted modules.

(d) Ceiling guard — autobot-backend/tests/test_raw_client_session_ceiling_12992.py does not exist on this branch's base (origin/Dev_new_gui @ 352e07cc7); it ships in still-open PR #12996, which currently has 10 red required checks (SAST, api-wiring, code-quality, smoke-test, etc.) and is not close to merging. There is therefore no ceiling constant in this PR to lower. Using that test's own AST-walker logic (copied, not modified) as a standalone script for measurement only: 87 on origin/Dev_new_gui @ 352e07cc7 → 72 on this branch, exactly −15 (the 15 sites converted here). Whoever finishes #12996 will re-measure against whatever tip it rebases onto, per that file's own "re-measure immediately before push" rule — this PR does not need to (and cannot, since the file isn't part of its diff) touch it.

Model Used

Claude Sonnet 5

[github-actions]

✅ SSOT Configuration Compliance: Passing

🎉 No hardcoded values detected that have SSOT config equivalents!