v0.8.3

Release v0.8.3

Proofreader accuracy overhaul.

Introduces a new audit-then-propose enforcement layer for the proofreader:

• Every proofreading batch now goes through submit_batch_audit before
  any changes are proposed. The model must reproduce each line's full
  content verbatim and mark a per-line verdict; propose_changes is
  gated to refuse anything targeting an unaudited or 'clean' line.
• Prompt drives sentence-by-sentence scanning with sentence-tagged
  notes (s1:, s2:, ...) so multi-issue lines and skipped sentences
  become visible. Explicit statement that a line may have any number
  of issues, not just one.
• Strict-scope clamp prevents stylistic drift: the proofreader now
  refuses to propose any change that can't cite a mechanical rule
  ('reads better' is disqualified by definition).

Model & batch changes:

• Default proofreader model bumped from Haiku (small) to Sonnet
  (medium) — Sonnet holds the multi-category attention required by
  the audit much more reliably.
• Proofreading batches capped at 25 lines regardless of tier so the
  model can genuinely re-scan each sentence.

Robustness:

• Audit content match now tolerates typographic variants (straight
  vs curly quotes, em/en dashes vs hyphens, ellipsis vs three dots)
  using the same normalization the proposal system already applies —
  mixed-typography manuscripts no longer trigger retry storms.
• Auto-reset behavior from v0.8.2 is preserved: re-proofreading a
  completed file works transparently.

21 new tests cover the audit tool acceptance path, every rejection
condition, quote/dash equivalence, and cross-file audit isolation.
Total suite: 914 tests, all passing.

Full Changelog: v0.8.2...v0.8.3