fix: DAG signatures treat i2's NotSet default as no default

[thorwhalen]

What

Part of step 2 of the re-land plan in i2mint/i2#48: meshed treats i2's NotSet sentinel, when it appears as a func node param's default, as "no default" while it builds a DAG's signature.

• New meshed.util.not_set_to_empty(param): param with no default if its default is NotSet.
• DAG.src_name_params applies it before merging same-named params and before sort_params.

Why

The independent review of i2mint/dagapp#4 tried meshed with a NotSet-defaulted node (what a re-landed i2mint/i2#88 FuncFactory shows) and found:

• Silent, high: DAG([g, h]).sig became (z, a=NotSet) instead of (a, z), because params with a default are sorted after required ones. dag(1, 2) would then bind 1 to z.
• Loud, medium: a NotSet param next to a same-named required param raised ValidationError: ... didn't have the same default in conservative_parameter_merge, so DAGs that build fine today would stop building.

With this PR the DAG signature is exactly what it is today, and a NotSet root is required, as it is today. FuncNode signatures themselves are untouched.

No-op against the current i2: no current signature carries NotSet.

i2 compatibility (decision)

Uses i2.is_not_set (i2 >= 0.1.74, i2mint/i2#91) when available, otherwise falls back on i2.deco.NotSet (present since 2023). The i2 requirement is unchanged.

Tests

meshed/tests/test_not_set_defaults.py: the DAG signature and positional call match the plain DAG's, merging with a same-named required param works, a NotSet root is required, and real defaults are kept. 3 of the 4 tests fail without the fix. There is also a doctest on not_set_to_empty.

Gate

• pytest --doctest-modules meshed (py3.12): 175 passed, 1 failed, 1 skipped. Master: 170 passed and the same 1 failure, which is pre-existing: test_hybrid_dag needs extrude/http2py, not installed here.
• Dependents run against this branch vs released meshed, identical results: front 72 passed, lookbook 132 passed, titbit 24, extrude 3, dagapp 2, allude 1. uf (39 failed / 66 passed), guided (1 failed / 1 passed), dotsci (2 collection errors), theremin (4 errors): same pre-existing failures on both. know, raglab-app, smart-cv could not be tested here (not on this box, or pyaudio needs system portaudio).
• Hosted CI: see checks.

🤖 Generated with Claude Code

[thorwhalen]

Independent refute-review (sub-agent): no blocker. It is a strict no-op for DAGs without NotSet (identity check only). DAG.partial (also with _consider_defaulted_arguments_as_bound), __getitem__ sub-DAGs, code_to_dag and parametrized_dag_factory all rebuild the DAG, so they get the fix. It also tried a real i2.FuncFactory with NotSet defaults by hand: the DAG signature is the same as today's.

Known limitation, documented rather than fixed here: the fix is at the DAG-signature level. The FuncNodes keep their NotSet defaults, so calls that bypass the DAG signature fill a missing root with NotSet where today they raise TypeError. These are FuncNode.call_on_scope, DAG.call_on_scope (and call_on_scope_iteratively, launch_debugger), and Slabs. For a FuncFactory node this is actually the factory's intended "call with nothing" behaviour, since it filters NotSet. A NotSet default can only reach any other node if that function declares it itself. Whether FuncNodes should also normalize is a step-3 decision and is recorded in i2mint/i2#48. dot_digraph still labels such a root a= (cosmetic).

test_real_defaults_are_kept is a no-regression guard: it is meant to pass with or without the fix.