test: regression coverage for code review fixes (CR-01/02/03, WR-04/05/07/08)

[W00DSRULES]

Adds 27 regression tests for fixes shipped in #12 that landed without test coverage. Test count: 123 → 150.

What's covered

tests/test_api.py (new, 8 tests) — first tests for apps/api/main.py

• Submit/get/list/delete endpoint happy paths + 404s
• CR-03 — `DELETE /tasks/{id}` cancels the in-flight `asyncio.Task` via the `_task_handles` dict
• CR-03 secondary — `_run_task` tolerates the `_tasks` entry vanishing mid-flight without `KeyError`
• WR-07 — `_tasks` `OrderedDict` is capped at `_MAX_TASKS` via LRU eviction at submit time

`tests/test_blackboard.py` (new, 17 tests) — direct Blackboard coverage

• CR-01 — `memory_guidance` is a valid entry type. Parametrized snapshot test over all 11 documented types so future renames/removals are deliberate
• WR-08 — Timestamps round-trip through `datetime.fromisoformat` with non-None tzinfo (catches naive `utcnow().isoformat() + 'Z'` regressions)
• WR-04 supporting — Per-session prefix isolation works correctly under a shared Redis instance
• Invalid type raises `ValueError`, `get_entries_by_type` filters correctly, etc.

Extension to `tests/test_core.py`

• CR-02 — `test_concurrent_executions_do_not_corrupt_self_blackboard`: runs two `execute()` calls concurrently on a shared `ThesisOrchestrator` and asserts `orch.blackboard` reference + `session_id` are unchanged. Verified that re-introducing the pre-fix `self.blackboard = Blackboard(session_id=session_id)` assignment makes this test fail.

Extension to `tests/test_budget.py`

• WR-05 — `test_unified_estimate_cost_includes_input_tokens`: asserts `_estimate_cost` reflects long prompt + system tokens, not just response length.

Verification

• 150 tests pass (`pytest tests/`)
• ruff E/W/F/I clean, black clean
• CR-02 test confirmed to fail when the pre-fix is re-introduced (regression-detection verified)

🤖 Generated with Claude Code