Feat: runtime reduction mode API (on/off/deterministic)

[itay-nakash]

Summary

Add a process-wide reduction mode settable at runtime via POST /labcx/mode, so an external signal (e.g. a vLLM KV-cache-pressure monitor) can activate/deactivate context engineering live — without restarting the proxy.

• Mode enum on|off|deterministic; ParseMode validates (unknown -> HTTP 400). off = transparent passthrough; on = reduce with the configured method; deterministic is reserved and currently behaves like on.
• GET /labcx/mode reports the current mode; POST/PUT sets it from {"mode":"..."} or ?mode=. Held in an atomic.Value.
• Reduce gate becomes: !bypassed(r) && mode != off. The per-request x-labcx-bypass header still forces passthrough. Default mode = on, so existing behavior is unchanged. Fail-open preserved.
• --mode initial-state flag on lab-cx proxy.
• TestRuntimeModeToggle: on/off/deterministic flips reduce<->passthrough, bad value -> 400, GET reports state.

Related issue(s)

(Optional) Testing Instructions

• make test (full suite, with the race detector) is green; go vet and gofmt are clean.
• Manual: start lab-cx proxy, then
  • curl -XPOST localhost:8080/labcx/mode -d '{"mode":"off"}' -> next reduce-eligible request is forwarded verbatim;
  • curl -XPOST localhost:8080/labcx/mode -d '{"mode":"on"}' -> reduction resumes;
  • curl -XPOST localhost:8080/labcx/mode -d '{"mode":"bogus"}' -> HTTP 400;
  • curl localhost:8080/labcx/mode -> {"mode":"..."}; /ready unaffected.

Fixes #