[everton-dgn]

Repeated unclosed subscripts followed by a final ] make recovery scan the remaining input again for each command. This caches bracket results after a failed search, only at positions the lexer visited, so recovery can still reinterpret quotes and comments correctly.

For the 256-line case in the issue, character reads fall from 204,539 to 10,737. The regression checks growth in reads rather than elapsed time and preserves the later-assignment checks.

The full 2,060-test suite passed on Node 24. After the final syntax adjustment, 263 focused tests, type checks, lint, formatting and the runtime smoke test passed. I couldn't run the Node 14 smoke test on this arm64 Mac; CI still needs to cover it.

Fixes #14

[jecruz]

Thanks for the fix, @everton-dgn. I reviewed this against #14 and the results look good.

What I tested (head b071427 vs its merge-base 61b82c2, which is current main; Node 24, built with plain tsc -p tsconfig.build.json):

1. Operation counts for Quadratic parsing of unclosed subscripts followed by a trailing ] #14. I counted String.prototype.charCodeAt calls during parse of 'a[x y\n'.repeat(n) + ']', n = 64..2048 (doubling), plus the 'a[$x y\n' variant.
   • Before: 14,267 reads at n=64 up to 12,646,395 at n=2048 (37 to 1,029 reads per character, doubling with n, i.e. quadratic).
   • After: 2,673 up to 86,001 reads, a constant ~7.0 reads per character. The $x variant is a constant ~8.1 (before: 46.5 to 1,322).
   • Without the closing ] the counts are identical before and after (5.0 and 5.6 reads per character), also linear.
2. AST differential. I serialized full ASTs for both builds and compared them byte for byte: 5,258 rows (policy probes, 4,096 seeded fuzz inputs, scaling up to 16 KiB) and 4,121 further rows, plus 8,192 seeded cases built from subscript, quote, comment, $(, ${, [[, (( and += atoms and 20,000 longer cases biased toward subscripts, quotes and here-docs. 0 differences and 0 exceptions on either side.
3. Recovery semantics. I checked the example from the issue (a[$x y # "\nb[$x z]=1\n# "), the single-quote variant, variants with an unmatched ] inside $(...), here-doc variants, an array body and a nested subscript. The ASTs are identical before and after, and the b[$x z]=1 assignment is still parsed as an assignment to b with index $x z.
4. Test suite on the PR head: all non-benchmark tests pass (2050/2050, including the new ones). test/benchmark.test.ts fails the same 8 tests before and after in my setup because I did not install the benchmark dependencies, so I treat that as unrelated.

I did not find a concern. This is differential evidence over the corpora above, not a proof of equivalence. One reading note: the per-start-position cache is only filled from positions the search actually visited and only when no errors were recorded, which matches the comment about not caching inside skipped quote/comment syntax. The 28,192 targeted cases above did not find a case where that breaks down.

[everton-dgn]

Thanks for the review and for the differential runs against #14. Let me know if anything else is needed before merging.