Match transformForLint masking to ember-estree's toPlaceholderJS (fixes per-file program rebuilds)

[NullVoxPopuli-ai-agent]

Fixes #229 (the performance/memory half, with a strong lead on the rest).

Root cause

For type-aware linting, every .gts/.gjs file gets a JS placeholder text twice, by two different implementations:

• Lint time: ember-estree's toPlaceholderJS keeps template content and blanks only `/$ (since ember-estree 0.6.10, the typescript-eslint error on a file #230 fix)
• Program build time: transformForLint in the patched ts.sys.readFile blanked every non-newline character (since v0.14.1, the Parser breaks on a specific template #226 fix)

typescript-estree hashes the code passed to parseForESLint against what the watch program last read from disk (getWatchProgramsForProjects). Since the two maskings differ, every file with a non-empty template hashes differently — which fires the file-changed watch callback and silently rebuilds the entire TS program inside getProgram(), once per .gts/.gjs file linted. It's invisible in debug logs ("Found existing program for file" prints either way); the tell is that services.program is a new object on every parse.

On a large codebase that's an O(project) rebuild × thousands of files — hour-long runs — and the allocation churn (× --concurrency workers) accounts for the tens-of-GB memory. It also plausibly feeds the non-deterministic type errors, since checker state now depends on per-worker lint order.

How the two fixes diverged

#226 and #230 were the same escape-overflow bug in the two copies of this logic. Through v0.14.0 both sides escaped `/$ identically, so bytes matched except on overflow files. The fixes chose different maskings (blank-all vs blank-`$), so as of v0.14.1/v0.14.2 every templated file mismatches.

Fix

Blank only `/$ in maskTemplateContentForLint, byte-matching ember-estree 0.6.10. This remains overflow-safe for the #226 case (1:1 replacement, placeholder never outgrows the region).

Verified against a minimal 3-file typed-lint project: before, services.program was a distinct object after every parse; after, all parses share one program.

Regression test

tests/placeholder-parity.test.js asserts transformForLint(code).output is byte-identical to the toTree placeholder across the #226 snippet, dollar-heavy expression templates, emoji/CJK, CRLF, and multi-template modules — so the two implementations can't silently drift again. Longer-term it may be worth exporting toPlaceholderJS from ember-estree and deleting the copy here.

🤖 Generated with Claude Code

[NullVoxPopuli]

good news is that all the other tests added over the last few days still pass