Changes in version 0.4.0                        

  - Introduced Registry v2 with separate model, provider, interface, and
    capability domains.
  - Refactored llm_caller() into a resolver, request builder, transport,
    response parser, and result-normalization pipeline without changing
    its experiment-facing behavior.
  - Added reusable protocol support for OpenAI-compatible Chat
    Completions, OpenAI Responses, DeepSeek-compatible chat, and
    Anthropic Messages.
  - Added explicit support levels and opt-in production Registry smoke
    tests. OpenAI Responses, DeepSeek Chat/Responses, and Qwen
    Chat/Responses passed live non-stream and stream verification for
    the documented model entries.
  - Preserved Registry v1 user files and the existing public registry
    return structures through compatibility projections.
  - Added network-free contract tests for request construction,
    streaming, response parsing, provider errors, timeouts, secret
    redaction, and bundled Registry v1/v2 equivalence.
  - Added adaptive_feedback_experiment() as the canonical
    adaptive-conversation entry point while preserving
    conversation_experiment_with_feedback() as a backward-compatible
    alias.
  - Corrected experiment result-path handling so an explicit output path
    no longer initializes the default user result directory.
  - Moved automatically named result files to the platform-specific
    package data directory returned by tools::R_user_dir() for
    CRAN-compliant persistence.
  - Added opt-in live DeepSeek coverage for ordinary conversation
    history and bounded adaptive feedback execution.
  - Moved the default user Registry to tools::R_user_dir() while
    retaining read-only compatibility with the historical ~/.psylingllm
    location.
  - Hardened Registry and result persistence so malformed Registry files
    are preserved, v1 registration cannot corrupt v2 bundles, parent
    directories are created explicitly, and failed writes no longer
    report false success.
  - Restored one-to-one RAW evidence retention across trial, factorial,
    conversation, adaptive-feedback, and multi-model experiment results.
    CSV and ordinary Excel result tables remain flat. The optional
    raw_output_path writes a lossless, redacted RDS archive; Excel
    output then adds compact RawIndex and RawMetadata sheets without
    embedding provider payloads in cells. Existing explicit saveRDS()
    workflows remain supported.
  - Added get_trial_raw(), extract_process_metrics(), and
    extract_token_probabilities() for provider-neutral inspection and
    analysis of supported reasoning-token, usage, and token-logprob
    evidence.
  - Changed omitted experiment-level generation_interface arguments to
    use the model's Registry default interface. Explicit interface
    values remain strict, public function signatures retain their
    historical displayed defaults, and ambiguous Registry v1 models
    retain the historical Chat fallback.
  - Decode non-stream provider response bytes explicitly as UTF-8 before
    strict JSON parsing. The existing control-byte recovery remains a
    secondary, auditable fallback, and other malformed JSON remains a
    structured parser error.
  - Added a narrow, auditable recovery path for otherwise valid provider
    JSON containing unescaped control bytes inside strings. Original
    response bytes remain available in RAW diagnostics; other malformed
    JSON remains a structured parser error.
  - Kept standard result columns present when their values are
    unavailable so streaming and non-streaming result files retain a
    stable tabular schema.