test: fix dev test-suite drift + harness bugs (no product changes)

[vybe]

Summary

A full-suite run on dev (Docker, local) surfaced 38 failures + 2 errors. Every one was root-caused to a stale test or a test-harness issue — zero product regressions. This PR is test-files-only (0 lines of production code).

Verified after fixes: 74 unit + 162 integration = 236 passed across the touched files.

Fixes (what drifted, why)

Cluster	Fix	Root cause
test_platform_prompt_unit (5)	stub lambdas accept **_kwargs	production compose path now forwards runtime= (#1187)
test_cb_probe_execution_close (7)	add get_setting_value to _CanaryTempDB	canary temp-DB stub leaks onto module-level from database import db; under full-suite order it serviced get_platform_default_model and AttributeError'd
test_fork_to_own (2 errors)	defensive list_all_agents_fast backfill	leaked bare services.docker_service stub (same sys.modules cross-file leak class)
test_files_guardrail_bypass, test_mcp_validator_endpoint (2)	assert "Disallowed credential file path"	intentional message wording change (#1305)
test_event_subscriptions (1)	expect 403 not 400	intentional #186 enumeration-safety collapse (commit 34ddf71a)
test_activities (1)	derive valid triggers from canonical _TRIGGER_BUCKETS	hard-coded allow-set drifted behind webhook / channel triggers; now tied to source of truth
test_whatsapp_integration (4)	read pending-login Redis via backend's authed client	bare redis-cli GET fails NOAUTH (Redis ACL mandatory since #589) but exits 0, defeating the returncode guard. Product code verified correct.

Clusters 1–3 pass in isolation and only failed under full-suite ordering (sys.modules leakage). A durable follow-up would harden the tests/unit/conftest.py sys.modules baseline restore — out of scope here.

⚠️ Not in this diff (environment, not code)

Running the full suite locally still shows a handful of failures/skips that are not fixed by this PR because they are environmental — a clean CI run is unaffected:

• 9 unit failures (bcrypt 72-byte error ×3, ModuleNotFoundError: PIL ×6) — a drifted local tests/.venv (bcrypt 5.0.0, no Pillow). requirements-test.txt already pins bcrypt>=4.2.0,<5 and Pillow>=11.1.0. Resync your local venv: cd tests && .venv/bin/pip install -r requirements-test.txt.
• ~60 infra-broken skips ("Agent server not ready", "testfix container not found", 429s) — the shared test agent was never ready (TEST_AGENT_NAME unset, no testfix fixture). Needs a persistent test agent.
• 4 quota failures (test_dynamic_thinking_status async ×3, one 429) — the real Claude subscription hit its usage cap during the 50-min run.
• 3 environmental (workspace_available/auto_switch defaults, agent_rename name collision) — pass in a clean CI env / fresh DB; code defaults are correct.

Test plan

• pytest on all touched files (leak-order preserved for canary→cb_probe): 236 passed
• Confirmed diff touches no src/backend/ production code

🤖 Generated with Claude Code

[dolho]

Review: ✅ Approve — safe to merge

Test-files-only, zero production lines (verified: diff touches only tests/). Each of the 40 failures is credibly root-caused to test drift or a harness bug, not a regression:

• test_platform_prompt_unit — stubs now accept **_kwargs to match the runtime= forwarding added in Codex harness MVP — pluggable agentic execution engine alongside Claude Code #1187. Correct.
• test_event_subscriptions 400→403 — matches the SEC: User and Agent Enumeration via differential API responses #186 enumeration-safety collapse (source-not-found folded into the not-permitted 403). Correct.
• *_credential file path message asserts — tracks the intentional feat(credentials): curated credential file-type injection (SA keys, certs, SSH, binary) #1305 wording.
• test_activities deriving valid triggers from _TRIGGER_BUCKETS — good move, ties the assertion to the source of truth instead of a literal that keeps drifting.
• test_whatsapp_integration reading pending-login through the backend's authed Redis client — real harness bug (bare redis-cli returns NOAUTH with exit 0 since SEC: Redis has no requirepass — agent containers have unauthenticated full read/write access to platform task queue (AISEC-C1) #589, silently defeating the returncode guard). Product code untouched.

CI green across all 3 seeds × base/head. One note (already called out in the PR): clusters 1–3 are band-aided per-file against sys.modules cross-file leakage; the durable fix is a conftest.py baseline restore. Fine to defer, but worth a follow-up issue so this doesn't recur file-by-file.