fix(tests): re-baseline raw-ClientSession ceiling to 13, document new SSRF-pinned carve-out, wire guard into CI (#13041)

[mrveiss]

Thinking Path

test_raw_client_session_ceiling_12992.py was failing on origin/Dev_new_gui (13 raw aiohttp.ClientSession(...) sites vs. a recorded ceiling of 12). Re-measured with the test's own AST walker (never grep -c — it over/under-counts on string literals) against a tree freshly synced to origin/Dev_new_gui at a2f788889: confirmed 13, same offender list #13041 reported.

Investigated each of #13041's three claimed "new, undocumented" offenders individually before touching anything:

• agent_loop/search/config_declared_provider.py — genuinely new (added by PR feat(research): topic-based source routing + data-driven search sources (#12625) #13016 / Research agent P3: topic-based source routing + data-driven source definitions #12625) and genuinely undocumented. Read the file: its raw session pins a TCPConnector via autobot_shared.security.ssrf_guard.pinned_connector() fresh on every call (config_declared_provider.py:183-187,235) — the same SSRF-pinned-connector shape already carved out for oauth_flow.py / external_importer.py. Pooling it would silently drop the pin and reopen the security(codeql): Server-side request forgery (SSRF) — 4 open code-scanning alert(s) #12278 DNS-rebinding hole. This site must stay raw; it is a legitimate carve-out, not a violation.
• orchestration/dag_executor.py and services/slm_client.py — checked git history of the test file: both were already listed in the module docstring's "custom non-default TLS/SSL context" bullet since batch 8 landed the 12-ceiling (perf(http): pool remaining raw aiohttp.ClientSession sites onto shared client (#12979) #13006). fix(http): raw aiohttp.ClientSession ceiling regressed 12→13 (undocumented offenders, #12992) #13041's claim that these two are new/undocumented is incorrect — only config_declared_provider.py is new.

So the fix is: raise the ceiling to the measured, correct value (13) and document the one genuinely new carve-out — not convert config_declared_provider.py to the pooled client, which would be a security regression.

Root cause of why this regressed unnoticed: the guard has never been collected by any CI workflow. autobot-backend/tests/** only appears in migration-gate.yml (migrations subdir only) and startup-import-smoke.yml (only test_startup_imports.py); the required "Unit & Integration Tests" check (frontend-test.yml) is entirely frontend. That gap is tracked separately as #10691 (evidence already posted there) — out of scope for a "smallest sound fix" PR to broaden. Within this PR's blast radius, the guard has zero non-stdlib dependencies (ast/pathlib only, no backend-module imports) and startup-import-smoke.yml's job already installs the full backend dependency set and is a required check — the natural, low-risk place to add exactly this one test.

What Changed

• autobot-backend/tests/test_raw_client_session_ceiling_12992.py: MAX_RAW_CLIENT_SESSIONS 12 → 13; module docstring's "Remaining raw sites" inventory updated to add agent_loop/search/config_declared_provider.py under the SSRF-pinned-connector category with its reasoning and refs (Research agent P3: topic-based source routing + data-driven source definitions #12625/feat(research): topic-based source routing + data-driven search sources (#12625) #13016/fix(http): raw aiohttp.ClientSession ceiling regressed 12→13 (undocumented offenders, #12992) #13041); MAX_RAW_CLIENT_SESSIONS's re-measurement history comment and the module docstring both record the fix(http): raw aiohttp.ClientSession ceiling regressed 12→13 (undocumented offenders, #12992) #13041 investigation, including which of its claimed offenders were already documented.
• .github/workflows/startup-import-smoke.yml: added a second step to the existing startup-import-smoke job (which already installs the full backend dep set and runs backend pytest, and is a required check) that runs test_raw_client_session_ceiling_12992.py, so a future regression fails CI instead of silently landing again.
• Did not touch config_declared_provider.py, dag_executor.py, or slm_client.py — all three stay raw by design.
• Did not attempt to broaden backend test coverage in CI beyond this one guard — that is ci: backend unit tests not enforced — stale assertion merged via #10623 undetected #10691's scope.

Verification

• Re-measured with the AST walker against origin/Dev_new_gui @ a2f788889: 13 constructions, offender list matches fix(http): raw aiohttp.ClientSession ceiling regressed 12→13 (undocumented offenders, #12992) #13041 exactly.
• pytest autobot-backend/tests/test_raw_client_session_ceiling_12992.py -v: 2 passed (walker-sanity + ceiling).
• Negative case: temporarily set MAX_RAW_CLIENT_SESSIONS back to 12 in the working tree and re-ran — reproduced fix(http): raw aiohttp.ClientSession ceiling regressed 12→13 (undocumented offenders, #12992) #13041's exact AssertionError: 13 raw ... exceeding ... 12 failure with the same offender list, then reverted to 13. Confirms the ratchet trips at ceiling+1.
• py_compile / flake8 --max-line-length=120 / black --check --line-length=120 on the changed test file: all clean.
• CI-wiring evidence: built a Python 3.14 venv (matches the workflow's actions/setup-python version), installed requirements-ci.txt + pytest + editable autobot_shared (the exact steps the job runs), then ran pytest autobot-backend/tests/test_startup_imports.py autobot-backend/tests/test_raw_client_session_ceiling_12992.py together and standalone — the ceiling guard collects and passes cleanly in both cases (2 passed), confirming the new workflow step actually executes the test rather than repeating the "silently never collected" failure mode fix(http): raw aiohttp.ClientSession ceiling regressed 12→13 (undocumented offenders, #12992) #13041 is about.
• Startup-import smoke ratio: unchanged — the new step is a separate, independent pytest invocation appended after the existing test_startup_imports.py step; test_startup_imports.py itself was not modified (350/350 passing locally against available deps, same as before this change).

Model Used

Sonnet

[github-actions]

✅ SSOT Configuration Compliance: Passing

🎉 No hardcoded values detected that have SSOT config equivalents!