Implement AST-based dependency extraction in FormulaVisitor.calculateDependencies (markdown-template)

[skypank-coder]

Summary

FormulaVisitor.calculateDependencies(tsCode) in packages/markdown-template/src/FormulaVisitor.ts is a stub (// TODO!!) that always returns []. It's invoked during TemplateMark typing (templatemarkutil.ts → templateMarkTypingGen), which sets dependencies on every FormulaDefinition / ConditionalDefinition, and the value is serialized through markdown-html (ToHtmlStringVisitor.ts writes the dependencies attribute, rules.ts reads it back). Today that attribute is always an empty array.

Continuing the working discussion with @mttrbrts

Current state / safety

Traced consumers of the field across the org:

• markdown-transform: only written (HTML attr), read back (rules.ts), or stripped (markdown-cicero/ToCommonMarkVisitor.ts deletes it).
• template-engine: only TemplateMarkInterpreter.ts does delete context.dependencies.

No code branches on the values, so implementing real extraction is self-contained with no regression surface — it only populates a field that is currently always empty.

Proposed approach

Extract root identifiers from the formula's TypeScript source (the param is named tsCode) via the AST rather than regex:

• Collect root identifiers that reference template variables / model, excluding locally-bound names, function params, and built-ins.
• Handle optional chaining (a?.b?.c → root a).
• Return [] on parse failure or unrecognized constructs (note: this changes the current catch, which throws — flagging so it's an explicit decision).

Open questions to scope before implementation

1. Runtime parser: typescript is currently a devDependency of markdown-template. Preference for promoting the TS compiler API to a runtime dep vs. a lighter parser (e.g. @typescript-eslint/typescript-estree)?
2. Dependency definition: which identifier classes count — model/variable refs only, and are locally-declared bindings and function params always excluded?
3. Cross-formula references: Devanshi noted formulas lack stable user-facing names, so these look out-of-scope for v1 — agreed to defer?

I can't make the Wednesday call, so opening this for async discussion per your
suggestion. Once (1) the runtime parser and (2) the dependency-definition
question are settled here, I'll scope the PR to those decisions.

[skypank-coder]

@mttrbrts Please go through this if all okay then ready to move further .