CHORE: Replace parser.ParseDir with golang.org/x/tools/go/packages

[philpennock]

parser.ParseDir is deprecated since Go 1.25. Use packages.Load with NeedTypes and NeedTypesInfo to correctly extract CanUse* constant integer values via TypesInfo.Defs, preserving the test's ability to detect missing providerCapabilityChecks entries.

---

This is a replacement PR for #4154. I took my review feedback from that PR, copied the markdown into two local files, and prompted Claude Code (using Sonnet 4.6) with:

Branch sx-pr/4154 contains one commit, which purports to "Replace parser.ParseDir, which is deprecated". But it is broken. This is a test of the codebase, via AST checks, that various providers.CanUse* checks are called appropriately. Files local.pdp.review1.1 and local.pdp.review1.2 (git-ignored) contain my review feedback, pointing out how the change is broken and how it can be validated. You are on a new git branch based off current main; please provide a correct replacement, and validate that it works.

The next section is choice quotes from Claude's analysis while working, and the final section was Claude's summary at end-of-work.

I have independently reviewed the code, it looks good, and I've validated that the tests look reasonable and correctly fail when a test is commented out (to ensure the AI did not fabricate that). I think this is a sane way forward.

I thought it would be amusing to use Claude to critique Claude's earlier work.

---

Now I understand the problem clearly. The broken PR sets capabilityInts[name.Name] = 0 for all constants (the comment even admits it), which means every CanUse* constant maps to int 0, and since capIntsToNames[0] finds the first entry, all checks pass trivially. The test is completely neutered.

The fix is to use packages.NeedTypes/NeedTypesInfo (already requested in the config) to actually read the integer values of the constants. Let me look at the providers package to understand the constants:

[...]

The problem is clear. The broken PR:

1. Uses packages.Load correctly to get type info (NeedTypes, NeedTypesInfo)
2. But then abandons the type info and sets all capabilityInts to 0, making every CanUse* constant map to CanAutoDNSSEC (which is in capIntsToNames[0]), so all checks pass trivially

The fix is to use providers.TypesInfo.Defs to look up the actual *types.Const value for each constant. Let me write the corrected test file:

---

Both validations from the review pass:

1. Correct output: integer values are real ((4), (9), etc.) and each is paired with its correct record type — matches the "before" output exactly.
2. Test actually catches omissions: commenting out the DNAME entry causes FAIL with MISSING: providers.CanUseDNAME (9).

What was wrong in the PR: The code switched from parser.ParseDir to packages.Load correctly, but the author gave up on extracting real integer values and wrote capabilityInts[name.Name] = 0 for everything. Since all constants got value 0, they all matched capIntsToNames[0] = AUTODNSSEC, making the test trivially pass regardless of what was in providerCapabilityChecks.

The fix: Instead of iterating the raw AST and guessing values, use providers.TypesInfo.Defs (which is what NeedTypesInfo provides) to look up each identifier's *types.Const object and call constant.Int64Val(c.Val()) to get the actual iota integer.