M6e: recursive tuple-stream detection (focus/index under sort/predicate)

[flearc]

Summary

Root-cause fix (via systematic debugging) for the last sort-on-focus joins failures. The tuple-mode detection was non-recursive: a ^(sort) or [predicate] nests a @$v/#$v step into a sub-path, so no top-level step carried .tuple and the whole path skipped eval_path_tuple entirely → tuple bindings lost → null.

• Fix (fe18f86), mirroring the already-recursive path_keeps_array:
  • path_is_tuple now recurses into nested path steps.
  • the path dispatch uses path_is_tuple(node) instead of a one-level scan.
  • the self-contained-step-1 seeding routes a nested tuple sub-path through eval_path_tuple(…, true) so its bindings survive (the gap M6c left in seeding; the main loop already did this).
• Baseline (038f860): regen, additions-only.

Results

• Official suite 1246 → 1249 (+3); joins 36 → 38/43 — employee-map-reduce/7,8 (sort-on-focus) plus a bonus sorting/case020 (a general correctness fix, not a 2-case patch).
• 491 unit tests green; zero-regression guard green (baseline diff additions-only; %-parent / sorting / transform unaffected).
• Oracle-faithful: case 7 byte-identical to jsonata-js v2.2.1; cases 7/8/case020 asserted against the official expected results; no-perturbation spot-checked.

Remaining joins gaps (deferred)

The reorder cases ($[[1..4]]#$pos[$pos>=2], $^($)#$pos[$pos<3] — filter/sort before #) still need jsonata's ordered-stages machinery.

Test plan

• busted spec/ — 491/0
• busted spec/jsonata_suite_spec.lua — zero-regression guard green
• bash scripts/run-suite.sh — 1249/1682, joins 38/43
• TDD: regression test fails without the fix, passes with it
• Oracle parity (jsonata@2.2.1) on the fixed cases

🤖 Generated with Claude Code