Define optional Japanese dictionary and default-off context-reading experiment #244

Problem

Japanese speech needs optional English loanword readings and better native
contextual readings. PR #237 introduces two public settings and resource IDs;
their compatibility and acceptance contract must be explicit before merge.
The imported CSV has independently reproduced shifted readings, and tsqyomi v5
has residual context errors and substantial process-memory cost.

Proposed public contract

• japanese-pronunciation: independently versioned optional CSV/provenance/license
  pack at assets/text/ja/english/. Official candidates admit only individually
  checked word/reading pairs. Unreviewed source rows remain local research data;
  source preservation does not authorize distribution. Native base-owned words
  retain base ownership. Keep whole-token boundaries and original source costs.
• ENABLE_JAPANESE_USER_DICTIONARY: allow installed official/personal/legacy
  dictionaries by default, preserving the existing local behavior. Missing packs
  do not install anything; disabling uses the base dictionary.
• japanese-context-reading: separate pinned local tsqyomi v5 model resource.
  It is a default-off experiment, with known context errors, not a default
  requirement or a generally validated Japanese reading solution.
• ENABLE_JAPANESE_CONTEXT_READING: default false, independent from the CSV
  setting. Installing model files does not enable it. Explicit activation may
  add roughly 650–900 MiB process memory in corpus testing; usage varies.
• No automatic model/CSV download; settings or resource changes require backend
  restart. Existing installer, native readiness and audio-cache identity contracts
  remain authoritative. Neither pack contains character voices or personal files.
• Older applications without the catalog/loader are incompatible with these new
  IDs. Publish a compatible application before distributing optional resources.

Acceptance

• Every retained official word/reading has an independent reviewed expectation,
  with all retained variants checked through production clean_text in context.
  Coverage-only lists cannot be described as correctness gates.
• Preserve original full sentences/digits for native regression checks. Include
  the external peer-review/neutral-faction/ball-count/stop-Work cases; distinguish
  existing native segmentation defects from model regressions and masked inputs.
• Report model memory, latency, known errors and default-off behavior in docs/UI.
  A failing experimental lexical gate remains a failing result, not an exception
  list added to runtime.
• Keep CPU/model-less use supported; retain third-party source/license records.
• Model/word data publication, human listening and independent maintainer review
  remain separate acceptance steps. feat(tts): add checked Japanese dictionary and default-off reading experiment #237 stays draft until those concerns resolve.

Scope and current work

Implementation/review: #237. The maintainer selected reviewed-only official
dictionary coverage, accepting a smaller pack. Unrelated HTTP dependency security
maintenance is isolated in #243. This issue records the proposed product contract;
it does not itself establish maintainer approval, completed listening or permission
to publish data/model archives.