Make harness startup and turns lighter

[justrach]

What changed

• Lazily materialize provider tool catalogs, Codex model catalogs, stored credentials, and other startup-only resources when they become observable.
• Batch stored credential reads and run independent macOS Keychain lookups concurrently.
• Run AI tab-title/header generation in a genuinely independent execution slot alongside the main model request, with the existing async fallback.
• Remove the 500 ms terminal cursor-query penalty in terminals that delay or drop DSR replies.
• Reuse context estimates, reset scratch allocations adaptively, and remove repeated serialization/allocation work from hot request paths.
• Parse JSON control envelopes with turn-local ownership and keep --json stdout strict JSONL, including fallback response text.
• Add long-session memory instrumentation plus live JSON and title-concurrency CI regressions; refresh generated SDK tool lists.

Why

Startup and lightweight REPL commands were paying for catalogs, credentials, schema rendering, and a terminal timeout before those resources were needed. Header generation could also serialize with the main turn. Long-lived JSON sessions retained transient envelope allocations.

Impact

• Boot-to-prompt median: 5.3–8.2 ms in controlled default/provider/exact-model runs.
• Lightweight command latency: roughly 500 ms down to 20–25 ms.
• Title and main request arrival: 0.3 ms apart in the final concurrency regression.
• Forty-turn JSON session-arena growth: 345 KB down to 200 KB (~42% lower); scratch remains flat at 9 KB.
• ReleaseFast binary: 3,176,200 bytes; --version smoke peaks around 1.82 MB RSS.

Durable conversation history and all provider/tool behavior remain intact; expensive data is deferred, not removed.

Validation

• zig build test --summary all — 215/215 tests passed
• Default and ReleaseFast builds
• JSON controls and strict live JSONL tests
• PTY REPL, Markdown, gutter, overflow, spinner, and model-preference tests
• Codex WS/SSE, compaction, stall/drop, and newer-issue E2E tests
• Title/main concurrency regression
• SDK Python/TypeScript validation and generation drift
• Zig formatting, source-size guard, and git diff --check