perf: fix all performance issues from issue list

[Copilot]

Fixes 15 performance-related issues found in the issue tracker.

TypeScript/Node.js fixes:

• perf(history-store): saveStore uses JSON.stringify with 2-space indent, bloating file size by 30–50% #77 – Remove JSON.stringify(…, null, 2) indentation in store.ts and settings-store.ts (machine-read files don't need pretty-printing)
• perf(agent-history): Math.max(...files.map(...)) spread in isDirty can stack-overflow with many sessions #76 – Replace Math.max(...files.map(f => f.mtime)) spread (stack-unsafe for large arrays) with .reduce() in agent-history/index.ts
• perf(input): stripReasoningContent calls JSON.stringify on the entire input array just to check field existence #73 – Replace JSON.stringify(input).includes('reasoning_content') with input.some(item => 'reasoning_content' in item) in gateway/bridge/input.ts to avoid full serialisation
• perf(usage-store): Math.max(...tsValues) spread overflows V8 argument stack on large record sets #71 – Replace Math.max(...tsValues) spread with a simple min/max loop in usage/store.ts
• perf(usage-store): computeUsageSummary re-reads pricing.json from disk on every call #70 – Cache mergedPricing() result to avoid re-reading pricing.json on every usage query; add invalidatePricingCache() for updates
• perf(thinking-state): ThinkingState.prune() uses O(n) Array.shift(), making each eviction linear-cost #74 – Replace O(n) Array.shift() with a head-index pointer in ThinkingState.prune() (gateway/deepseek/state.ts)
• perf(adapter): CodexDeepSeekAnthropicAdapter.conversations Map grows unbounded — missing LRU eviction #75 – Add evictOldest() (with while loop) to enforce LRU eviction on conversations and thinkingStates Maps in adapter.ts
• perf(session-cache): collectFiles() called again after dirty check, duplicating stat syscalls #69 – Avoid calling collectFiles() twice per dirty agent in session-cache.ts by returning { dirty, files } from isAgentDirty()
• perf(history-scanner): readFileSync reads the entire session file when only the first 150 lines are needed #68 – Replace full-file readFileSync + split with a synchronous 64 KB chunk read using openSync/readSync in all agent-history scanners; use StringDecoder to safely handle multi-byte UTF-8 at the chunk boundary

Rust fixes:

• perf(hover-tracker): element_at_point_for_hover serializes ElementInfo to JSON then parses it back to HoverElement #63 – Replace serde_json::to_value(&info) + parse_hover_element() round-trip in hover_tracker.rs with direct field mapping; re-export ElementInfo from each platform module
• perf(find-image): Template scaling uses Lanczos3 filter, 3–5× slower than Bilinear with no benefit for template matching #60 – Switch template image scaling from Lanczos3 to Triangle in find_image.rs (~4× faster with acceptable quality loss for template matching)
• perf(find-image): precompute_template allocates an unnecessary Vec<f64> intermediate buffer #61 – Remove intermediate Vec<f64> allocation in precompute_template with a two-pass (mean then norm) loop over the image pixels
• perf(find-image): Clones full screenshot (~15 MB Retina image) when no search region is specified #59 – Use Cow<GrayImage> in find_image to borrow the screenshot directly when no search region is specified, avoiding a full clone
• perf(screenshot-cache): ScreenshotCache::peek() uses VecDeque linear scan, O(n) lookup #58/perf(image-cache): ImageCache::get() uses VecDeque linear scan, O(n) lookup #57 – Replace linear-scan VecDeque<CachedScreenshot/CachedImage> with HashMap<id, data> + VecDeque<id> for O(1) lookups in ScreenshotCache and ImageCache
• perf(screenshot): Full-screen capture performs redundant JPEG→PNG round-trip, wasting 4 image operations #56 – Eliminate the extra spawn_blocking JPEG→PNG round-trip in handle_take_screenshot; both the JPEG (for the response) and PNG (for the cache) are now produced in a single blocking task