fix(analysis): count elements the way the client-checked OverPy counter does

[e54-bot]

Fixes #305.

Program::element_count was about 37% below the client on a production project (OWBastion/Bastion, zh-CN main): 18891 against the client's 30070 for the OverPy build, whose own counter says 30067. The model now reproduces OverPy's figure to within 2 elements.

What changed

Derived by comparing every action and condition of that project against pinned OverPy 9.7.10's #!debugElementCount (2867 leaf actions, 309 rules), then fitting the missing costs:

• a number literal costs two (the value and its literal); Team, Hero, Color, Button and Map constants likewise;
• a variable read costs two (the value and its name), a variable named as an argument counts as that argument;
• a comparison outside a rule condition costs two (its operator is a literal);
• an omitted defaulted argument still fills its slot (Custom String has three replacement slots);
• Else If, Else and End are actions; the count follows the canonical emitter, which closes the last action of a rule without End when it is an If and keeps every other End (a written and an omitted trailing End are one program);
• Workshop setting values get the client's fixed adjustments;
• hero pairs are counted per direct argument;
• disabled rules, actions and conditions count like enabled ones instead of failing the whole program.

Evidence

On the project's main entry every action count equals OverPy's and the total is 30065 against OverPy's 30067. Eight rules differ by one where the emitter's End spelling differs from OverPy's (documented in docs/element-count.md). The client counted OverPy's build at 30070; the three elements of difference are not explained. The model is calibrated on this one project; constructs it does not exercise follow the Workshop.codes reference. docs/element-count.md states this, and tests/element_count.rs reproduces ten small OverPy-counted programs, one rule each.

Verification

cargo fmt --check, cargo clippy --workspace --all-targets -- -D warnings, cargo test --workspace --all-targets, workshop-catalog-gen check, git diff --check. The full-project comparison was run locally against OverPy's annotated output; the project source is not part of this repository.

Counts for existing programs change, so consumers that pinned a total will see different numbers.

[Teakowa]

One blocking count-model issue.