Fix two live bugs reported in the PR backlog

[adamjohnwright]

Two defects that were still present on main, both reported by contributors. Neither contributor PR is merged as-is; see below.

Alliance MITAB columns — reported by @AaryanCode69 in #120

Adjacent string literals concatenate silently in Python, so a missing comma turns two column names into one. molecular_interaction had three such sites, collapsing seven columns into three:

"Interaction detection method(s) Publication 1st author(s)"   -> 2 columns
"Host organism(s)Interaction parameter(s)"                    -> 2 columns
"Annotation(s) interactor A" ... "Creation date"              -> 6 columns

Those metadata fields were never populated on Alliance documents.

genetic_interaction is the same MITAB schema written correctly, so it serves as an oracle: both lists are now 41 columns with identical sets.

#120 caught two of the three. The third — "Interaction detection method(s) Publication 1st author(s)" — was a single literal containing a space rather than an implicit concatenation, so it did not appear in that diff.

Why this is guarded by a test, not a lint rule

ruff's ISC001 flags implicit concatenation on one line, but it is disabled because ruff itself warns it conflicts with the formatter. And running ruff format joined "Host organism(s)" "Interaction parameter(s)" into a single literal, destroying the syntax a rule could match on. Two tests instead: one comparing the MITAB lists against each other, one rejecting any column name that begins with another. Both were run against the pre-fix source to confirm they fail on the real bug rather than passing vacuously.

AgentGraph.del — reported by @bleedblack1 in #156 and @bhavyakeerthi3 in #147

__del__ called asyncio.run() unconditionally, which raises RuntimeError when a loop is already running — and __del__ fires at arbitrary moments, including inside the running server. Exceptions there are swallowed and printed, so it appeared as log noise with the pool still open.

Neither proposed fix is taken as-is:

• fix: make AgentGraph destructor loop-safe to prevent connection leaks #147 schedules the close with loop.create_task(). A task created from __del__ is not guaranteed to run if the loop is shutting down — which is exactly when a graph is usually collected — so it reports success while leaking the pool.
• fix: remove asyncio.run() from __del__ to prevent production crashes #156 rewrites the module more broadly than the bug warrants.

This closes the pool only when no loop is running, and otherwise warns that it is still open rather than pretending otherwise. Errors from the close are caught, because __del__ must not raise.

This is containment, not a cure. Nothing calls close_pool() explicitly, so __del__ is the only cleanup that exists, and it is fundamentally unreliable — writing the test surfaced that even the warning fails at interpreter shutdown once logging's handlers are closed. The real fix is closing the pool from an application shutdown hook, which belongs with the agent-API work.

Tests

97 total, up from 91. Four cover the destructor: loop running, no loop, no pool, and a close that raises.

Closes #120. Supersedes #147 and #156.

🤖 Generated with Claude Code