Declare the input port, naming the paths the Jinja templates ask for

[seebi]

Closes the last finding from the review of #12: the task declared no input_ports at all,
so what the workflow editor offered on the input side was DataIntegration's default rather
than anything the task stated.

What it declares now

One port when the query or the variables carry Jinja syntax — which is exactly when the
task reads its input — and no port at all otherwise, since a task without Jinja sends
its text once and ignores whatever is connected. A handle there only invites a connection
that does nothing.

The port names the paths, taken from the templates themselves with
jinja2.meta.find_undeclared_variables, the way cmem-plugin-llm derives an input schema
from its template variables. This is the part that matters: DataIntegration reads only the
paths a consuming task requests, so an unknown or flexible input asks for nothing and is
handed nothing.

A run that takes the per entity branch and receives nothing now warns in the execution
report, instead of reporting a successful run having sent no query at all. That was noted
as out of scope on the last ticket; it is also what made the investigation below possible,
since the warning is what reported how many input collections arrived.

What this does not fix

A dataset wired into this task still delivers zero entities, and the declaration is not
the reason. Verified against a deployment, five ways, all reporting Received 0 input
collection(s): no declaration, UnknownSchemaPort, FixedSchemaPort with the derived
paths, FlexibleSchemaPort, a freshly created workflow, and a distinctly versioned build
after a forced uninstall and reinstall.

What is established:

The source dataset yields entities	yes — 2 rows, attribute id
The workflow wiring	identical in shape to a transform reading a dataset in a workflow that works
Operator into this task	input is delivered — chaining a second task in feeds the entity through
Dataset into this task	nothing arrives, in every variation above

So the remaining suspect is the dataset to Python task schema negotiation, or something
about workflows created through the API rather than drawn in the editor. That wants either
a look from someone who knows the DataIntegration side, or a check of whether the editor
behaves differently — not another guess here.

Tests

Three new tests: the derived input schema names the template variables, a task without
Jinja declares no input, and a task with Jinja requests the paths it renders. task check
green — 38 offline, ruff, mypy, deptry, trivy.

🤖 Generated with Claude Code

[github-actions]

Coverage Report

File	Stmts	Miss	Cover	Missing
init.py	0	0	100%
workflow/init.py	0	0	100%
workflow/graphql.py	149	27	82%	242-243 273-275 277-278 280 289-294 316 323-325 328 346 355-360 416
workflow/utils.py	102	0	100%
TOTAL	251	27	89%

Tests	Skipped	Failures	Errors	Time
50	13 💤	0 ❌	0 🔥	3.151 ⏱