perf(core): optimize agent hot path across 9 subsystems

[bug-ops]

Summary

Resolves #621 — optimize the core agent loop by eliminating unnecessary allocations, O(n) scans, and blocking operations across 9 targeted subsystems.

• Enable SQLite WAL mode + SYNCHRONOUS=NORMAL for 2-5x write throughput (Enable SQLite WAL mode and SYNCHRONOUS=NORMAL #639)
• Replace O(n*iterations) token scan with cached_prompt_tokens (Replace O(n*iterations) token scan with incremental counter #640)
• Defer maybe_redact to stream completion boundary (Defer maybe_redact to flush_chunks() boundary #641)
• Replace format_tool_output allocation with Write-into-buffer (Replace format_tool_output allocation with Write-into-buffer #642)
• Change ToolCall.params from HashMap to serde_json::Map (Use serde_json::Map directly for tool call params #643)
• Pre-join static system prompt sections into LazyLock<String> (Pre-join static system prompt sections into LazyLock #644)
• Replace doom-loop string history with content hash comparison (Replace doom-loop string storage with content hash #645)
• Return &'static str from detect_image_mime with case-insensitive matching (Return &'static str from detect_image_mime #646)
• Replace block_on in history persist with fire-and-forget async spawn (Replace block_on in history persist with spawn_blocking #647)

10 files changed, +198 -84, 6 new tests (2017 total passing).

Test plan

• cargo +nightly fmt --check passes
• cargo clippy --workspace -- -D warnings passes (0 warnings)
• cargo nextest run --workspace --lib --bins — 2017 passed, 9 skipped
• WAL journal_mode assertion with tempfile-based DB
• doom-loop hash equality/inequality tests on Vec<u64> path
• detect_image_mime mixed-case extension tests
• format_tool_output structure tests