test: add 12 integration tests covering identified coverage gaps

[markphelps]

Summary

• Adds 12 new integration tests covering gaps identified by auditing the existing 125-test suite
• Enhances the test harness to support error body assertions and webhook error message capture

Harness Enhancements

• ! curl now writes the HTTP response body to stderr, enabling tests to assert on error response content (e.g., 409 bodies)
• webhookResult captures the actual error_message string (not just has_error bool), enabling assertion on failed webhook payloads

New Tests

Test	Gap Covered
prediction_error_response	JSON response shape when predict() raises ValueError/RuntimeError/Exception
nested_output_types	BaseModel-in-BaseModel, List[BaseModel], Optional nested fields
iterator_error_midstream	Generator yields partial output then raises — webhook reports failure
predict_sys_exit	sys.exit() fails prediction but worker survives for next request
TestConcurrentAboveLimit	max_concurrency + 1 returns 409 Conflict with capacity error
sequential_state_leak	Documents module-level state persistence across sequential predictions
webhook_delivery_failure	Server stays healthy when webhook URL is unreachable
TestSIGTERMDuringSetup	SIGTERM during slow setup() causes clean shutdown
webhook_prediction_error	Failed prediction webhook payload includes error message
build_openapi_schema_complex	Schema for constraints, choices, Secret, Optional, structured output
coglet_large_file_upload_serial	50 MiB binary file through upload URL pipeline
async_generator_precollect	Async generator output is pre-collected into response array

Test Types

• 10 txtar tests in integration-tests/tests/ (declarative testscript format)
• 2 Go-native tests in integration-tests/concurrent/ (require parallel HTTP coordination)

How to Run

mise run build:cog && mise run build:sdk && mise run build:coglet:wheel:linux-x64
COG_BINARY=dist/go/*/cog mise run test:integration

[markphelps]

Code Review Fixes

Pushed 1f1531ce addressing the review feedback. Here's what changed and what I intentionally left as-is:

Fixed

#	Issue	Fix
1	TestSIGTERMDuringSetup sends SIGINT, not SIGTERM	Changed os.Interrupt → syscall.SIGTERM
2	TestConcurrentAboveLimit 500ms race window	Replaced fixed sleep with polling loop that retries the overflow request until 409 (10s timeout)
3	TestSIGTERMDuringSetup 3s race window	Replaced fixed sleep with waitForServerStatus() polling /health-check for STARTING status (60s timeout)
7	Weak predict_time assertion	Changed from key-existence check to regex verifying a decimal number value
8	Regex . as quote wildcard	Added comment explaining this is intentional — testscript single-quoted args cannot embed literal single quotes, so . is the pragmatic choice

Not changing

#	Issue	Reasoning
4	webhook_delivery_failure exec sleep 5	The real assertion is the subsequent sync prediction succeeding, not the sleep. No reliable poll target exists without knowledge of coglet's internal backoff schedule.
5	! curl stderr has no consumer	Infrastructure code that enables a test pattern. Reasonable to land capability ahead of tests that use it.
6	predict_sys_exit worker identity	Distinguishing "same worker survived" from "worker respawned" would require exposing worker PID through the predictor/IPC protocol. The observable behavior (second prediction succeeds) is what the test cares about.
9	Missing [short] skip	Checked the broader codebase: most cog serve tests (healthcheck_*, coglet_metrics, coglet_large_input, etc.) do not use [short] skip. Only notably slow tests (GPU builds, large files, subprocess tests) use it. The new tests are consistent with the majority pattern.

[michaeldwan]

looks good, thanks for the fixes!