test(agent-eval): add output-style A/B harness, cost/token analyzer, and DISALLOW/REP_START controls

Three additions to tighten the eval loop:

- offload-eval-styles.sh: new 4-arm eval (raw/refs/map/src) isolating the Worker's
  output shape's effect on main-session tokens, latency, and accuracy. Delegation
  blocked by default (DISALLOW=Agent) so variance from Haiku subagent spawning doesn't
  contaminate the measurement.
- offload-eval-cost.mjs: cost/token analyzer that reads Claude Code's own per-model
  accounting (modelUsage.costUSD) rather than re-deriving from raw token counts,
  giving a correct main(Sonnet)/sub(Haiku) split with proper per-tier pricing.
- offload-eval-3arm.sh: adds DISALLOW env to block sub-agent delegation across all
  arms, and REP_START to append reps to an existing run without clobbering earlier
  jsonls (e.g. REP_START=4 REPS=3 → reps 4,5,6).

Also adds CODEGRAPH_OFFLOAD_STYLE forwarding to the managed gateway so the styles
eval can drive output shape end-to-end; the field is stripped before the upstream
model call and never sent to BYO endpoints.

src/reasoning/reasoner.ts

@@ -200,6 +200,9 @@ export async function synthesizeOffload({ query, context }: SynthArgs): Promise<
   const url = cfg.url.replace(/\/+$/, '') + '/chat/completions';
   const { system, footer } = promptFor(cfg.style);
   const ctx = cfg.strip ? stripAgentDirectives(context) : context;
+  // Optional operator/eval flag forwarded verbatim to the managed Worker (see body below);
+  // the Worker validates it and falls back to its default for anything it doesn't recognize.
+  const workerStyle = (process.env.CODEGRAPH_OFFLOAD_STYLE || '').trim();
 
   const controller = new AbortController();
   const timer = setTimeout(() => controller.abort(), cfg.timeoutMs);
@@ -217,6 +220,10 @@ export async function synthesizeOffload({ query, context }: SynthArgs): Promise<
         max_tokens: cfg.maxTokens,
         temperature: 0.2,
         reasoning_effort: cfg.effort,
+        // Optional managed-tier flag, forwarded ONLY to the managed gateway (which strips it
+        // before the upstream model call) and ONLY when an operator/eval sets it — so BYO
+        // endpoints, which may reject unknown fields, never see it.
+        ...(cfg.managed && workerStyle ? { offload_style: workerStyle } : {}),
         messages: [
           { role: 'system', content: system },
           {