feat(ab-corpus): bump minimax max_tokens to 8192 (defensive)

ab-v4 raw output capture on dashboard-search-filters-composite shows
minimax-m2.7 emitting <think>...</think> + 4 op_tool tags that fit
inside the 4096 default — its measured completion-token average for
this run was 697, well under the cap. So thinking-budget truncation
is NOT the actual root cause of minimax's lower multi-tool hit rate
(25% vs gpt+deepseek 50%); the real issues are instruction-following
(mixed Strategy A + B despite the explicit forbidance, invented
"canvas" parent_id placeholder).

Still doubling the cap defensively: composite multi-tool outputs can
chain 12-13 op_tool tags + thinking, and "fit easy" today doesn't
mean "fits headroom-free on a longer brief tomorrow." The bump is
free on the happy path (provider stops generating when done, doesn't
bill unused headroom) and only ever helps when the model would
otherwise hit a real ceiling.

Real follow-up for minimax: instruction compliance — the no-mix rule
needs to land harder than a single trailing sentence. Probably wants
the rule moved to top-of-prompt + a few-shot bad-example contrast.
Out of scope here; tracked under Phase 2 prompt design.

scripts/ab-corpus/clients/minimax.ts

@@ -33,7 +33,16 @@ export async function callMinimax(args: CallMinimaxArgs): Promise<ChatCallResult
     system: args.system,
     user: args.user,
     temperature: args.temperature,
-    maxTokens: args.maxTokens,
+    // M2.7 emits chain-of-thought inside the same `content` channel
+    // (we strip <think> blocks in output-parser.ts); the openai-compat
+    // default 4096 is fine for obvious prompts but cuts close on
+    // composite multi-tool outputs (12-tag responses + thinking can
+    // approach 3-4k easy). Double the cap defensively — we already pay
+    // for thinking either way, and most replies still come in well
+    // under 1k completion tokens (ab-v4 avg 697), so the bigger cap
+    // costs nothing on the happy path and only matters when the model
+    // would otherwise truncate mid-output.
+    maxTokens: args.maxTokens ?? 8192,
     label: 'minimax',
     // ab-v3 saw 4 minimax timeouts on 104 runs — all wall-clock aborts,
     // none were content-quality issues. One retry picks up the