fix: reliability batch 0A — secure-field capture, durable dictionary writes, language-correct filler cleanup

[GalaxyRuler]

Before Submitting This PR

Please confirm you have done the following:

• I have searched existing issues and pull requests (including closed ones) to ensure this isn't a duplicate
• I have read CONTRIBUTING.md

Human Written Description

I ran a deep three-round adversarial audit of the dictionary, context-awareness, and language pillars and this is the first containment batch from it: seven bounded reliability fixes, no new features. The ones I care most about are the privacy fix (selected-text transforms could read password fields on all three platforms) and the two that silently corrupt bilingual usage — filler cleanup keyed to the UI language instead of the dictation language, and Android applying dictionary entries the user had quarantined.

Related Issues/Discussions

Fixes # — internal reliability audit findings (no public issue filed)
Discussion: —

What's in the batch

1. Selected-text transforms fail closed on secure fields (Win/mac/Linux) — password/secure check runs before any text read; error while checking = secure. New stable reason codes secure_field / secure_check_error, localized toast on the shortcut path.
2. Adaptive reprocess blocked in private sessions — gate is the first side-effecting decision (private_session_active).
3. UTF-8-safe punctuation extraction — byte-index panic on Arabic/emoji fixed via char_indices.
4. Android formatter integrity — snapshot now serializes only active dictionary entries; startup no longer overwrites the on-device snapshot with empty/failed store loads.
5. Ambiguous dictionary-ID mutations rejected — duplicate persisted IDs can no longer cause delete-one-deletes-all (ambiguous_entry_id); migration comes in the next batch.
6. Durable dictionary writes — new fallible locked+forced-save settings writer; commands and the post-paste learner report success and emit events only after the value is persisted (rollback on save failure).
7. Filler cleanup keyed to validated dictation language — English filler lists no longer applied to Arabic/other dictation; auto/multilingual modes skip language-default fillers; explicit custom fillers always apply.

Testing

• Full Rust suite: 727 passed, 0 failed (run twice — implementation session and an independent review session).
• tsc, ESLint, Prettier, cargo fmt --check, translation completeness (869 keys × 19 locales), Bun unit tests for the Android snapshot module: all pass.
• Windows device verification: probe against a live UIA password box returns secure_field before any text read; normal editable field still captures. mac/Linux and Android on-device checks are listed as follow-up debt.

Screenshots/Videos (if applicable)

—

AI Assistance

• No AI was used in this PR
• AI was used (please describe below)

If AI was used:

• Tools used: Codex (implementation), Claude Code (audit, adversarial review rounds, code review, device verification)
• How extensively: implementation and review were AI-driven end to end from a human-approved hardening plan; every item was code-reviewed against spec and the security item was device-verified on Windows before push

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c53f45510e

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Use the output language for translated cleanup

When translate_to_english is enabled, result.text is already English, but this call still passes effective_language derived from the selected source language (or None in auto mode) into filter_transcription_output and the adaptive cleanup path. In translated dictation from auto or any locked non-English language, default English filler removal is skipped, so the new language-correct cleanup regresses translated output; derive the cleanup language from result.translated/the translation target before applying local transforms.

Useful? React with 👍 / 👎.