chunkshop 0.9.1

Fixes a regression introduced by 0.9.0's path-less language detection (#69):
generated and minified files that 0.8.3 skipped were being symbol-parsed,
flooding downstream consumers with chunks and OOM-ing them (pg-raggraph#79,
bento). Python-only; the Rust crate is a lockstep version bump with no
functional change.

Fixed

• symbol_aware no longer over-parses generated / minified files (#71).
  0.9.0's content heuristic started classifying machine-emitted files as code
  that 0.8.3 skipped. A 143 KB generated .ts (3,000 trivial functions) became
  3,000 symbol chunks; consumers that embed every chunk OOM'd. Two complementary
  guards:

  • Content-detection guard (path-less only).
    detect_language_from_content now returns None for files that look
    machine-emitted — an @generated / sourceMappingURL marker, or a minified
    (very long, >2000-char) line — so they fall back to sentence_aware
    (bounded) instead of being symbol-parsed. Explicit signals (cfg.language,
    metadata['language'], a real path) bypass the guard, so a caller can
    still force such a file through.
  • Per-file symbol-chunk cap. When a document would emit more than
    max_symbols_per_file symbol chunks, the chunker logs a warning and falls
    back to sentence_aware with fallback_reason="too_many_symbols". Catches
    pathological generated files regardless of how the language was resolved.

  Under defaults, the 2,500-function generated .ts now yields 62 bounded
  chunks (was 2,500) and a 46 KB minified one-liner yields 23; normal code is
  untouched.

Added

• SymbolAwareChunker.max_symbols_per_file (default 2000, null to
  disable, must be >= 1) — caps symbol chunks per document. The default
  catches generated files while leaving even very large hand-written sources
  alone (real code rarely exceeds a few hundred top-level symbols).