[performance] Preserve useful highlight checkpoints across distant jumps

[owainlewis]

Goal

Keep syntax-context work bounded when a user alternates between distant locations in a large file.

Evidence

At exact main 1d6d60189e4b350649f591a858746368de78064c, WindowedHighlightCache::push_checkpoint in src/highlighter.rs retains only checkpoints before the new line and then removes old nonzero checkpoints beyond the recent cap. After warming a deep viewport, a sufficiently large backward jump can rescan from line 0 and delete the useful later checkpoints. Jumping forward again repeats depth-proportional work.

Original review thread: #119 (comment)

Acceptance criteria

• Alternating among distant warmed viewports performs work bounded independently of file depth after bounded warm-up.
• The checkpoint set has a deterministic memory cap and preserves strategically useful context on both sides of a jump.
• Edits invalidate only checkpoints whose syntax context may be stale.
• Rust comments and strings plus Markdown fences retain correct context.

Checks

• Add a deterministic regression that alternates between locations farther apart than the current recent-checkpoint window and asserts scanned lines and request bytes stay bounded.
• Cover an edit between the distant locations.
• Run focused highlighter tests and opt-in highlighting performance checks.
• Run cargo fmt --check, cargo test --all-targets, strict Clippy, and an all-target build.

Scope

Keep the fix inside the bounded checkpoint/window design. Do not add background parsing or a new parser framework.

[owainlewis]

Implemented in ready PR #144: #144

Commit: b0040db on exact base b3ef75d.

Verification: 44 focused highlighter tests passed; all 4 opt-in performance checks passed; cargo test --all-targets passed with 308 tests plus all 5 signal cleanup tests; formatting, diff validation, strict all-feature Clippy, and all-target build passed. Fresh independent exact-base review approved with no findings. Protected CI is pending.