megahal-cli-v0.3.1

Fixed

• (symbol) align MegaHalSymbol::new with upper() using to_ascii_uppercase (#9)
• (gen) preserve input order for keyword/seed scan to match C reference
• (compat_v8) load dictionary words as raw bytes, no UTF-8 validation (#10)
• (gen) seed a keyword even when the forward root is childless
• (gen) restructure generation loop to do/while, matching C reference
• (tokenizer) return zero tokens for empty input, matching C

Other

• describe loader behavior instead of the prior code in the v8 raw-bytes test

megahal-cli-v0.3.0

Added

• add C MegaHAL v8 .brn import (#3)

Fixed

• compatibility with rand 0.10.1

megahal-cli-v0.2.0

First release on crates.io. Renames the inner bidirectional Markov crate from markov-chain to megahal-markov to avoid a name conflict, exposes new APIs in the facade, and fixes four algorithm bugs relative to the C reference.

Added

• MegaHal::save_brain_to_writer / load_brain_from_reader for
  in-memory brain serialization without filesystem I/O.
• MegaHal::train_from_reader for training from any BufRead source.
• MegaHal::generate: a no-learn variant of respond.
• MegaHal::set_fallback_reply / set_fallback_greeting for replacing
  the hardcoded "I don't know enough..." and "Hello!" messages.
• MegaHalError for reader and writer errors.
• megahal-cli crate with cargo install megahal-cli for the megahal
  binary, extracted from the workspace root. cargo binstall megahal-cli
  also works, pulling the prebuilt binary from the matching GitHub release.
• Runnable examples under crates/megahal/examples: chatbot (stdin
  loop), game_hud (no-learn generate with a tight iteration budget),
  and brain_persist (save/load via reader/writer).

Changed

• Renamed the markov-chain crate to megahal-markov. The name
  markov-chain is held on crates.io by an unrelated 2017 crate.
• SymbolId's tuple-struct field is no longer public; construct via
  SymbolId::new(value). The symbol-core crate's API is documented as
  unstable; consumers should depend on the megahal facade.
• Expanded the megahal crate-level rustdoc with an API tour, concepts
  section, brain-file format note, and a recipe for shipping a
  pre-trained brain via include_bytes!.
• Running megahal without --brain now defaults to persisting the brain
  to the user's data directory instead of running with no persistence.
• MegaHal::train_from_file now streams the file line-by-line, meaning
  a mid-file UTF-8 decoding error leaves the model partially trained
  instead of validating UTF-8 upfront.

Fixed

• capitalize no longer upper-cases the letter after a !/./? when
  no whitespace follows (e.g. a.b.c now correctly produces A.b.c).
• Reply generation no longer echoes the user's input when that was the
  only thing the model could produce; falls back to the canned reply
  instead.
• The backward generation phase now runs even when the forward phase
  produces no tokens, matching the C reference behavior.
• babble no longer panics when a context node's usage count is zero;
  treats that case as sentence-terminating.

v0.1.0

0.1.0