Finish two more fixes that merged incomplete

[dovvnloading]

Problem

Both shipped earlier today with the defect still reachable.

Budget caps could still restore unusable — #393

That commit claimed the plan's caps "keep falling back to their real defaults rather than zero, since a restored plan with max_steps=0 could never run again". They did not. _non_negative_int is max(0, int(value)) with a fallback only when int() raises, so every numeric route to zero went straight through — and the outer or default only catches Python-falsy values, so the string "0" and "-5" passed it too:

max_steps=-5    -> 0        max_steps='0'  -> 0
max_steps='-5'  -> 0        max_steps=0.5  -> 0
max_steps=0     -> 12       (the one case the test covered)

The harness restorer was worse — its local _int helper passed negatives through unchanged:

max_turns=-5    -> -5       max_context_tokens=-1 -> -1
max_turns='0'   ->  0

builder._spend_breach rejects the node on its first tick, so it lands on the canvas and can never run again.

The field-comment gate could not see the modules it was written to cover — #400

That PR corrected content's comment to "TEN kinds" and added a gate to pin it. Both were wrong. backend/domain/groups.py:330 and :368 build kind="frame" and kind="container" nodes with content= — twelve kinds, not ten. And the gate hard-coded graph.py and session_load.py while its own docstring called them "the two modules that create nodes", so a thirteenth kind landing in groups.py would have left it green.

That is the exact failure the gate exists to prevent — a closed set asserted by hand, growing, nothing failing — reproduced inside the gate itself.

Change

_positive_int is the right helper for a cap, where zero is not a value; both restorers use it. Counters keep _non_negative_int, because 0 is a real count. The harness's local _int is gone in favour of the two shared helpers.

The gate now discovers every module under backend/ that constructs a SceneNode instead of naming them, with a test that fails if discovery ever narrows back to the original two. The kind sets stay hand-authored on purpose: a human deciding "yes, a thirteenth kind should write content" is the checkpoint.

Test plan

• 9 parametrised plan-cap cases and 6 harness-cap cases covering every numeric and non-numeric route to zero or negative, plus one pinning that a real cap survives and a counter of 0 is not rewritten to a default.
• A discovery test asserting the module scan finds groups.py alongside the original two.
• tests/test_node_state_migration.py caught the first draft of these tests aliasing node.state — rewritten to the required X.state.<field> chain rather than weakening the gate.
• Full suite: 3181 passed, 19 skipped. ruff check . clean.

🤖 Generated with Claude Code