TEST: Stabilize target startup and analytics deadline tests

[hannahwestra25]

Description

Merged current main (ba20617d5eaddf8ae302d7f32454296ba50a9cc7) and narrowed this PR after #3278 landed. The diff is now five files instead of 21. It preserves #3278's completion-wait and analytics-budget fixes, #3221's frozen analytics-identity clocks, and #3255's deterministic SQLite contention checks.

Post-#3278 CI still exposed two target-test timing failures and one separate analytics assertion failure:

Evidence	Targeted change
macOS Python 3.13 merge-group job: test_cancelled_send_preserves_cancellation_when_retirement_fails_async exceeded its two-second wait for the mocked send to start.	Use the existing finite 300-second deadlock-backstop convention for that startup wait only. Keep the separate two-second cancellation-completion assertion unchanged.
macOS Python 3.14 merge-group job: the discovery provider-state test's success variant reported "Awaited 0 times" with a five-second probe deadline that also covers preceding memory access.	Give this test a finite 300-second correctness backstop. For its timeout variant, wait until the mocked request is pending, then expire it through the real asyncio.wait_for cancellation path. Assert cancellation occurred and drain the owned task on every exit.
Windows Python 3.12 post-merge job: the event-controlled analytics lifecycle test failed on assert 300.0000000000002 <= 300.	Freeze both deadline clocks in that test and assert the exact 300-second budget. Expiration still comes from its explicit event, with the existing HTTP 504, busy-capacity, repeated-cancellation, and cleanup assertions intact.

The shared TEST_WATCHDOG_SECONDS names the existing 300-second backstop; wait_for_completion_async retains its existing default and behavior.

Removed the blanket watchdog increases, changes to unrelated correctness-query budgets, and the Makefile/workflow worker override. CI still uses four workers. The distinct five-second scorer-cancellation cleanup check is preserved.

No production code, dependencies, provider settings, retries, matrix policy, or JUnit/reporting behavior changes. The targeted startup waits remain finite, but a genuine startup deadlock can now take five minutes to fail. These changes do not establish that reducing workers would improve reliability or that every historical failure was a runner stall.

Tests and Documentation

Local validation on Windows with Python 3.14.4, after uv sync --frozen --extra github-copilot --link-mode=copy:

uv run --frozen --no-sync -m pytest -n 4 --dist=loadfile -q -p no:cacheprovider --tb=short tests\unit\backend\test_analytics_lifecycle.py tests\unit\analytics\test_execution.py tests\unit\backend\test_analytics_routes.py tests\unit\prompt_target\target\test_github_copilot_target.py tests\unit\prompt_target\test_discover_target_capabilities.py

Result: 319 passed, with 32 existing deprecation warnings. This includes the complete changed test modules, analytics deadline/cancellation contracts, and related HTTP route tests.

uv run --frozen --no-sync ruff check tests\unit\async_utils.py tests\unit\backend\test_analytics_lifecycle.py tests\unit\prompt_target\target\test_github_copilot_target.py tests\unit\prompt_target\test_discover_target_capabilities.py
uv run --frozen --no-sync ruff format --check tests\unit\async_utils.py tests\unit\backend\test_analytics_lifecycle.py tests\unit\prompt_target\target\test_github_copilot_target.py tests\unit\prompt_target\test_discover_target_capabilities.py
uv run --frozen --no-sync -m ty check --output-format concise tests\unit\async_utils.py tests\unit\backend\test_analytics_lifecycle.py tests\unit\prompt_target\target\test_github_copilot_target.py tests\unit\prompt_target\test_discover_target_capabilities.py
uv run --frozen --no-sync -m build_scripts.validate_docs
git diff --cached --check

Results: all passed; all four Python files were already formatted.

A temporary, uncommitted pytest plugin also checked the actual baseline and revised test bodies under identical bounded conditions: a three-second mocked Copilot startup, six-second mocked discovery startup, and a frozen clock where (268.72848822480245 + 300) - 268.72848822480245 rounds above 300. The five baseline cases intentionally failed with the expected timeout, zero-await, or budget-assertion signatures; all five revised cases passed. No real providers or live workflows were invoked.

doc/contributing/4_running_tests.md describes the targeted watchdogs and frozen deadline clocks. JupyText is not applicable because no notebooks or executable documentation samples changed. The full unit suite, all-extras environment, and macOS execution were not run locally.

[Roman Lutz (romanlutz)]

hannahwestra25 I took the liberty to update this after #3279 merged. The more of these fixes go in and the sooner the better.