test(agentic): add 46 state management and orchestrator tests — all passed

[Mog9]

Continues the agentic loop test coverage from #1. This PR adds tests for the state management layer (agent_state.py) and the main orchestrator loop (agent_runner.py), plus the remaining 5 tests for tool dispatch (agent_tools.py).

46 new tests across 3 files — all passing.

test_agent_tools.py (5 new tests)

#	Test	What it validates
1	test_unknown_tool_returns_error	Unknown tool name returns error ToolResult
2	test_tool_exception_returns_error	Handler raising exception is caught, returns error
3	test_run_verify_delegates_to_verifier	Calls verify_candidate with correct args
4	test_run_benchmark_delegates_to_benchmarker	Calls benchmark_candidate
5	test_read_candidate_file_rejects_bad_path	Path traversal in read returns error

test_agent_state.py (16 new tests)

#	Test	What it validates
1	test_init_state_creates_file	agent_state.json written to disk
2	test_init_state_truncates_transcript	Old transcript cleared on fresh start
3	test_init_state_defaults	Status=PENDING, iteration=0, cost=0.0
4	test_save_and_load_roundtrip	State survives JSON serialization
5	test_load_returns_none_when_missing	No file returns None
6	test_load_returns_none_on_corrupt_json	Malformed file returns None
7	test_request_abort_sets_flag	abort_requested becomes True
8	test_request_abort_only_when_pending_or_running	Returns False for terminal statuses
9	test_request_abort_returns_false_when_no_state	No state file returns False
10	test_append_transcript_adds_timestamp	Each line has "at" key
11	test_append_transcript_multiple_lines	Multiple appends produce multiple lines
12	test_read_transcript_empty	No file returns []
13	test_read_transcript_skips_blank_lines	Blank lines ignored
14	test_read_transcript_skips_invalid_json	Malformed lines ignored
15	test_state_path_format	Path is repo_root / artifact_dir / "agent_state.json"
16	test_transcript_path_format	Path is repo_root / artifact_dir / "agent_transcript.jsonl"

test_agent_runner.py (25 new tests)

Cost tracking (6 tests)

#	Test	What it validates
1	test_cost_from_usage_all_zeros	Zero tokens = $0.00
2	test_cost_from_usage_input_only	1M input tokens = $5.00
3	test_cost_from_usage_output_only	1M output tokens = $25.00
4	test_cost_from_usage_cache_write	1M cache_write tokens = $6.25
5	test_cost_from_usage_cache_read	1M cache_read tokens = $0.50
6	test_cost_from_usage_combined	Mixed tokens = correct sum

System prompt (3 tests)

#	Test	What it validates
7	test_build_system_prompt_has_two_blocks	Rules block + cached skill bundle block
8	test_build_system_prompt_cache_control	Second block has cache_control: ephemeral
9	test_build_system_prompt_includes_task_details	Op, GPU, dtype, shape in rules text

Loop termination (7 tests)

#	Test	What it validates
10	test_loop_errored_on_missing_api_key	No ANTHROPIC_API_KEY = ERRORED immediately
11	test_loop_errored_on_api_refusal	stop_reason="refusal" = ERRORED
12	test_loop_breaks_on_end_turn_without_tools	end_turn with no tool_use = breaks
13	test_loop_rejected_on_give_up	Agent calls give_up = REJECTED
14	test_loop_rejected_on_max_iterations	Loop exhausts iterations = REJECTED
15	test_loop_rejected_on_cost_cap	Cost exceeds cap = REJECTED
16	test_loop_errored_on_api_error	APIStatusError = ERRORED

State transitions (5 tests)

#	Test	What it validates
17	test_iteration_counter_increments	State.iteration matches loop count
18	test_cost_accumulates_across_turns	cost_usd grows with each API call
19	test_token_counts_accumulate	All four token counters grow correctly
20	test_verify_result_tracked_in_state	last_verify_passed/reason updated
21	test_benchmark_result_tracked_in_state	last_benchmark_passed/speedup updated

Edge cases (4 tests)

#	Test	What it validates
22	test_loop_continues_when_verify_fails	Verify fails = no promotion, loop continues
23	test_loop_continues_when_benchmark_fails	Benchmark fails = no promotion, loop continues
24	test_promotion_failure_does_not_crash	promote_candidate raises ValueError = handled gracefully
25	test_transcript_written_each_turn	transcript_lines increments each turn

Results

test_agent_tools.py: 15 passed (10 from #1 + 5 new)
test_agent_state.py: 16 passed (new file)
test_agent_runner.py: 25 passed (new file)
Full suite: 149 passed, 1 skipped

Coverage summary

Module	Before	After
agent_tools.py	0%	15 tests
agent_state.py	0%	16 tests
agent_runner.py	0%	25 tests
Total agentic	0	56