fix: #88 add windows x86 setup - #108

The empty-tools wedge, the SOTA feature batches, and a hardening pass.

TRANSPORT (the "research tools return nothing" wedge): the SDK's
in-process MCP transport intermittently DROPS tool results — handlers run
and return content, the model sees "completed with no output", and the
agent reports a fake outage (upstream claude-agent-sdk-typescript marktext#108,
claude-code #43642/#36319; the same payloads survive over HTTP). WordBird
now serves its 57 tools from a loopback Streamable-HTTP MCP server
(agentSdk/httpMcpServer.ts); tools still execute in-process via
runForModel, so the proposal/snapshot/mode-gating pipeline is unchanged.
Gating moved off the canUseTool permission stream onto a matcher-scoped
PreToolUse hook, and every non-spawn SDK built-in is denied (AskUserQuestion
was collapsing the stream).

FAILURE RESILIENCE (load failures are expected; wedging is not):
- per-subagent maxTurns, and one-at-a-time spawning with a stale-slot
  reclaim so a subagent that dies without reporting cannot wedge the gate
  shut ("supervisor waits forever")
- _withSilenceTimeout: a runtime that goes completely silent ends the turn
  as a BUDGET event ("say continue"), per-message timer so slow-but-alive
  work is never cut short
- renderer stall watchdog threshold made injectable and finally covered
- 9 negative tests inject each failure and assert the reaction

SOTA batches: lore auto-injection, voice priming, anti-slop corpus lint,
craft fields, writing goals, relationship map, manuscript import,
scene-beat drafting.

AUDIT: fixed real bugs found while reviewing — manuscript import silently
dropped pre-chapter text; always:true bible pages could be crowded out by
the injection cap; imported projects lacked .gitignore/style.md; mermaid
labels were not escaped for `|`; a binder listener leak; the session
counter never started on a new project. Consolidated 4 copies of
escapeRegExp / 3 front-matter parsers / duplicated tokenizers into
markdownText.ts, and 12 e2e scaffold copies into fixtures.

Tests: 1223 unit, 177 e2e, lint + typecheck + build green.

Co-Authored-By: Claude Opus 4.8 <[email protected]>