fix(highlighter): reseed fenced Markdown segments

[owainlewis]

Summary

• Reseed each bounded Rust/Markdown highlight segment from the lexical context active at that segment's first line.
• Advance retained fence, quote-depth, and injected-Rust context through each segment while keeping every oversized line as a context barrier.
• Add focused unquoted and block-quoted regressions with a retained fence opener, oversized middle line, visible Rust suffix, real closer, following Markdown heading, parser-request count, and retained-work bounds.
• Record the scoped design and verification plan in docs/issues/146-plan.md.

Why

PR #144 retained Markdown fence context after a truncated line, but each post-barrier parser segment still started as a standalone Markdown document. Without a synthetic fence opener, visible Rust after the oversized line lost injected-language highlighting until a later window rebuild.

This change keeps the existing segmentation and checkpoint design. It gives each already-bounded segment the context Tree-sitter needs, then advances the same compact lexical state used by checkpoints before parsing the next segment.

Test plan

Verified on exact base 90f2823ded43d102650a8bce1e80e652fdb3740f and head 0ee87bb797539acef0dcb8afe56ea417d7d9db14:

• cargo test highlighter::tests:: -- --nocapture: 46 passed.
• cargo test performance:: -- --ignored --nocapture --test-threads=1: all 4 passed; deep highlighting completed in 1.281s initially and 25.104ms after edit.
• cargo test --all-targets: 317 unit tests passed, 4 documented performance tests ignored and run separately, and all 5 signal cleanup tests passed.
• cargo fmt --check.
• git diff --check origin/main...HEAD.
• cargo clippy --all-targets --all-features -- -D warnings.
• cargo build --all-targets.
• Applied the named improve workflow to clarify the segment-barrier invariant and preserve behavior.
• Fresh independent review inspected the final tracked and untracked diff, ran focused and all-target tests, and approved with no important findings.

Risks

The fix deliberately does not remove or relax truncated-line barriers. Each existing parser segment receives only the bounded synthetic wrapper for its retained context, and parser call count remains one per segment. Window lines, per-line source prefixes, checkpoints, scanned work, and the #124 exact Rust wrapper cap keep their existing limits.

Real fence closers are scanned and parsed inside their segment, so following Markdown resumes on the next line. Block-quoted fences keep the retained quote depth. The separate #124 overflow branches and #142 overflow recovery policy are unchanged.

Related issue

Closes #146

[owainlewis]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. 🎉

Reviewed commit: 2c5f69deb7

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".

[owainlewis]

Merged-main verification complete.

Merge commit: 2c5f69deb75109501e4d682c259de616098b45dc

Main CI: https://github.com/owainlewis/cortex/actions/runs/30407212734

Rust checks passed in 1m45s, including the security workflow policy, formatting, Clippy, tests, and release build. Issue #146 is closed, and the original PR #144 review thread is replied to and resolved with the merged fix and focused regression evidence.