fix(ai): corrected thinking config format and model context windows across 100+ definitions

- Fixed thinking configuration format by replacing `levels` array with `minLevel`/`maxLevel` properties across 100+ model definitions.
- Corrected GPT-5.4 mini/nano context window from 400000 to 272000 tokens for accurate token limit reporting.
- Normalized GPT-5.4 variant priority handling to use parsed variant instead of raw model IDs for consistent behavior.
- Added "mini" variant support to OpenAI model parsing regex and updated thinking mode configuration for Claude models.
- Fixed test robustness by replacing exact string matching with numeric range comparison to handle BSD seq notation on macOS.
- Corrected model generation script execution order to apply policy overrides before promotion target linking.

packages/coding-agent/test/bash-executor.test.ts

@@ -248,10 +248,15 @@ describe("executeBash", () => {
 		// Truncated output should be within the spill threshold
 		expect(result.outputBytes).toBeLessThanOrEqual(DEFAULT_MAX_BYTES);
 
-		// The tail of the output should contain numbers near the end of the range.
-		// The exact last number may be split across a truncation boundary, so
-		// check for a number within the last 1000 lines.
-		expect(result.output).toContain(String(lineCount - 500));
+		// The tail should still contain numeric values near the end of the range.
+		// BSD `seq` on macOS formats large numbers in scientific notation, so parse
+		// the final lines numerically instead of matching one exact decimal string.
+		const tailValues = result.output
+			.split("\n")
+			.slice(-1000)
+			.map(line => Number(line.trim()))
+			.filter(Number.isFinite);
+		expect(tailValues.some(value => value >= lineCount - 500 && value <= lineCount)).toBe(true);
 
 		// With 64KB read buffer, ~40MB should produce ~600 chunks, not 5M.
 		// Allow generous headroom but ensure it's orders of magnitude below lineCount.