Preserve shared lexical environments in compact EtsIR JSON

[CaelmBleidd]

The acquired lodash.js dependency expands shared closure-environment types into 755,694,861 JSON characters, exceeding Node's string limit and failing full-project frontend generation with RangeError: Invalid string length.

Serialize complete lexical environments once in a file-local table and use indexed references at their use sites. The Kotlin reader validates references and memoizes conversion to the existing environment model, preserving every captured local and function signature. Legacy inline JSON remains supported; invalid indices and cycles fail explicitly.

Validation:

• 49 focused frontend checks and TypeScript typecheck passed.
• 8 focused Kotlin reader/conversion checks passed, including legacy input, sharing, bounds, cycles and context cleanup.
• The unchanged frozen 569-file real graph generated all 569 JSON files. The offending file shrank to 12,424,070 bytes while preserving 52 classes, 693 methods, 617 environments and all 2,001 captured locals.
• Actual Kotlin decoding took 394 ms and model conversion 86 ms with -Xmx768m.

No symbolic machine or full test suite was run. Final combined frontend packaging and experiment startup remain separate validation steps.

Diff scope: production changes span 7 files (+179/-9 lines); focused tests and test-consumer helpers span 6 files (+372/-5 lines).

[github-actions]

Test Results

  248 files  +1    248 suites  +1   16m 16s ⏱️ + 1m 30s
  900 tests +8    888 ✅ +8  12 💤 ±0  0 ❌ ±0 
2 170 runs  +8  2 120 ✅ +8  50 💤 ±0  0 ❌ ±0 

Results for commit 199ebe0. ± Comparison against base commit 1260b68.

♻️ This comment has been updated with latest results.