perf: every cdp() call spawns a fresh Node process — expose the daemon client in-process

[apmantza]

Where: extractors/common.mjs:82-106 (cdp() spawns node bin/cdp.mjs per command), bin/cdp.mjs daemon client (~856-975)

The per-tab daemon architecture correctly reuses one CDP WebSocket per tab, but every command still pays a full Node CLI startup (~80-150ms on Windows). An all-engine search issues dozens of cdp(...) calls (list/nav/eval per engine, plus per-tick polls that remain node-side), so this is seconds of pure spawn overhead per search.

Fix: extract the daemon client (socket connect + sendCommand) into an importable module (e.g. src/search/cdp-client.mjs); have search.mjs, extractors, and fetch paths send commands over the existing named-pipe socket in-process, keeping bin/cdp.mjs as a thin CLI wrapper for manual use. Related micro-item: captureMainContext's unconditional 100ms settle sleep (bin/cdp.mjs ~362) could resolve early on the root frame's Runtime.executionContextCreated event once the client is in-process.

[apmantza]

Implementation attempt (by a plegma gpt-5.6-luna worker) was functionally correct but got reverted from the batch branch before merge. Findings:

• The refactor itself works: src/search/cdp-client.mjs daemon client, bin/cdp.mjs as thin wrapper with identical output/exit contract, in-process rewiring in extractors/common.mjs + src/search/chrome.mjs. Stealth-check passes 20/20; visible-mode searches work; 102/102 unit tests.
• But headless Perplexity regresses: interleaved A/B runs on the established profile showed Cloudflare challenging the in-process build while the spawn-per-command build passed, same cookies. On a zero-reputation scratch profile both builds get challenged, pointing at behavioral scoring: removing the ~100-150ms process-spawn gap between every CDP command makes the automation cadence faster/burstier, and CF notices.
• Suggested next attempt: keep the in-process client but add deliberate pacing — a small randomized minimum inter-command delay (e.g. 40-80ms jittered) on page-interacting commands (eval/click/type/nav), skipping it for list/evalraw bookkeeping calls. That preserves most of the win (spawn cost gone, bookkeeping fast) while restoring human-ish rhythm where CF watches.
• The attempt's code is preserved on the (reverted) commit c21029c on perf/batch-4 history / worker commit c5ce5d5.