v0.1.2 Conformance, sequence-scan fusion

[vlad-public-code]

Conformance work, the sequence-scan fusion performance pass, and a docs
trim, brought together onto main.

Changelog

Performance

• Sequence-scan fusion (translator/scan_fusion.py, new): groups a block's
  operations per bound sequence and emits one helper that reads each distinct
  field once, feeding every accumulator from that read. On the analytical
  benchmark: 551 field reads per evaluation down to 320, 1 955 function calls
  down to 1 253, evaluation 1.33x faster. It plans rather than rewrites —
  operations are collected only from unconditionally evaluated positions — and
  falls back to the original helper for anything but a plain int/float
  field.
• $seq[field = <literal>] compiles to a monomorphized filter_field_eq call
  instead of a per-element callback (hoisted lambdas on the benchmark: 18 → 3).
• Objects built from literal keys the translator has proved distinct skip
  object_of's per-key duplicate check.
• Cold compilation rises ~1.5x (~17 KB → ~29 KB of generated module), the
  deliberate cost of the above; it repays after ~100 evaluations.
• Current figures: 106 µs evaluation vs jsonatapy 285 µs, jsonata-rs
  517 µs, jsonata-python 5 797 µs.

Conformance

• Built-in context substitution: $fn() as a path step, and direct calls
  with fewer arguments than the signature, now route through a port of the
  reference parseSignature/validate logic; the evaluation context is
  substituted into focus (-) parameters by argument type at runtime. Covers
  45+ previously diverging cases ($power(3) crash, $uppercase() wrong error
  code, …). direct_call_is_safe() exempts fully-determined fast paths from the
  validation cost.
• Tuple paths (runtime/tuples.py, translator/tuple_path_codegen.py,
  both new): tuple-stream path evaluation and its codegen.
• Parenthesized path head: ("ab").foo is a value navigation (undefined),
  not a field lookup.
• Wildcard / descendant: ported reference evaluateWildcard /
  evaluateDescendants; leading-step outer-wrapper case separated from the
  per-element case via wildcard_context and at_input_root.
• Nested path wildcard fold: the optimizer no longer flattens an inner
  PathExpr containing a wildcard or descendant step into the outer path,
  preserving the re-split semantics of a nested evaluatePath.
• $spread: singleton-collapses the single-object result and preserves the
  array result, fixing the previous inversion.
• Datetime picture formatting/parsing, decimal and integer pictures, English
  number words, regex operations and string built-ins reworked against the
  reference.
• _leaks_evaluator_scope uses a whole-token regex, so call_builtin_ctx in
  hoisted closure bodies no longer false-fires.

Tests

New suites for scan fusion, tuple-path bindings, signature validation, callback
hoisting, object construction, array-constructor path heads, constructor step
flattening and value wrappers, path-stage/null conformance, reference
conformance, cache bounds, memory (generated modules stay collectible) and
concurrency.

Docs

• Performance section in README.md and docs/index.md cut by ~40% — every
  table, figure and caveat kept; the per-session measurement diary that had
  accumulated around them removed.
• Added a "Sibling implementations" table, and corrected the jsonata2js speedup
  to the figure that project reports for itself (~53x-60x vs jsonata).
• Version bumped to 0.1.1; docs/llms.txt refreshed.

🤖 Generated with Claude Code

https://claude.ai/code/session_015XC8GqzZHFbZzTw8pEjwVs

---

This discussion was created from the release v0.1.2 Conformance, sequence-scan fusion.